<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Siray AI on Medium]]></title>
        <description><![CDATA[Stories by Siray AI on Medium]]></description>
        <link>https://medium.com/@siary_ai?source=rss-81e778058314------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*8-JQU5QyMUm3SGtqNXeYSA.png</url>
            <title>Stories by Siray AI on Medium</title>
            <link>https://medium.com/@siary_ai?source=rss-81e778058314------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 12:26:14 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@siary_ai/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[GPT Image 2.6 Flare vs Sunburst: Speed and Quality Compared at the Same Price]]></title>
            <link>https://medium.com/@siary_ai/gpt-image-2-6-flare-vs-sunburst-speed-and-quality-compared-at-the-same-price-0d139d76e3d9?source=rss-81e778058314------2</link>
            <guid isPermaLink="false">https://medium.com/p/0d139d76e3d9</guid>
            <dc:creator><![CDATA[Siray AI]]></dc:creator>
            <pubDate>Fri, 20 Nov 2026 07:13:55 GMT</pubDate>
            <atom:updated>2026-11-20T07:13:55.108Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_609QbbaCLWhYfdJ.png" /></figure><p>GPT Image 2.6 Flare vs Sunburst is a choice between two OpenAI image models released on November 4, 2026: OpenAI positions Flare for fast iteration and Sunburst for detailed work and consistent edits. On Siray AI both cost exactly the same, $0.031 to $0.039 per medium-quality image at common sizes, so price doesn’t decide it. We ran three Black Friday product prompts and one edit on both.</p><h3>Key takeaways</h3><ul><li><strong>Same price, every cell.</strong> The Siray AI pricing tables for Flare and Sunburst are identical, from $0.015 (low, 768×1024) to $0.09 (high, 4K) per image, with a 10% discount at time of writing.</li><li><strong>Same parameters.</strong> Both take the same fields (size, quality, n, background, output_format including svg and gif, complexity).</li><li><strong>Speed was not what we expected.</strong> In our runs Sunburst finished 3 of 4 tasks faster (21 s vs 47–120 s). Flare was faster only on the flat lay (46 s vs 139 s). Tasks ran in parallel, so treat this as one sample, not a benchmark.</li><li><strong>Both rendered exact poster text.</strong> “WINTER SALE”, “Nov 27 — Dec 1” and “Northwind Candle Co.” came out spelled correctly on both.</li><li><strong>Both edits stayed in bounds.</strong> A “change only the color” edit kept shape, handle and shadow on both models.</li></ul><h3>What OpenAI says the difference is</h3><p>OpenAI introduced the two models together: “Flare generates images in as little as six seconds. Sunburst brings higher-quality generation and greater consistency across edits” (<a href="https://x.com/OpenAI/status/2118124964197114039">OpenAI on X</a>, November 2026). Both support animated GIF output, SVG export and a complexity setting. The Flare/Sunburst split started with GPT Image 2.5 in September; see our <a href="https://blog.siray.ai/gpt-image-2-5-api-pricing/">GPT Image 2.5 pricing guide</a>.</p><h3>Endpoints on Siray AI</h3><ul><li><strong>openai/gpt-image-2.6-flare-t2i</strong> — Task: Text to image; Main inputs: prompt, size</li><li><strong>openai/gpt-image-2.6-flare-edit</strong> — Task: Edit; Main inputs: prompt, images (1–16), optional mask</li><li><strong>openai/gpt-image-2.6-sunburst-t2i</strong> — Task: Text to image; Main inputs: prompt, size</li><li><strong>openai/gpt-image-2.6-sunburst-edit</strong> — Task: Edit; Main inputs: prompt, images (1–16), optional mask</li></ul><p>Sizes: 1024x1024, 1024x1536, 1536x1024, 768x1024, 1024x768, 2560x1440, 1440x2560, 3840x2160, 2160x3840, auto or custom with width and height. Quality: low, medium, high, xhigh, max.</p><h3>Pricing on Siray AI</h3><p>Both models bill by tokens; Siray AI shows a per-image conversion. These rows are identical for Flare and Sunburst (November 20, 2026, 10% discount applied):</p><ul><li><strong>low</strong> — 1024×1024: $0.022; 1024×1536 / 1536×1024: $0.018; 2560×1440: $0.023; 3840×2160: $0.042</li><li><strong>medium</strong> — 1024×1024: $0.039; 1024×1536 / 1536×1024: $0.031; 2560×1440: $0.043; 3840×2160: $0.077</li><li><strong>high</strong> — 1024×1024: $0.047; 1024×1536 / 1536×1024: $0.037; 2560×1440: $0.05; 3840×2160: $0.09</li></ul><p>A worked example: 100 vertical 1024×1536 product posters at medium quality cost about $3.10 on either model. Square images cost more than vertical or horizontal ones at the same quality. Edit endpoints use the same table, plus a small charge for input image tokens. Check live prices on the <a href="https://console.siray.ai/?ref=blog.siray.ai">Siray AI console</a>.</p><h3>Quickstart</h3><pre>curl -X POST https://api.siray.ai/v1/images/generations/async \<br>  -H &quot;Authorization: Bearer $SIRAY_API_KEY&quot; \<br>  -H &quot;Content-Type: application/json&quot; \<br>  -d &#39;{<br>    &quot;model&quot;: &quot;openai/gpt-image-2.6-sunburst-t2i&quot;,<br>    &quot;prompt&quot;: &quot;E-commerce product photo of a matte black ceramic pour-over coffee kettle ...&quot;,<br>    &quot;size&quot;: &quot;1024x1024&quot;,<br>    &quot;quality&quot;: &quot;medium&quot;,<br>    &quot;output_format&quot;: &quot;png&quot;,<br>    &quot;n&quot;: 1<br>  }&#39;</pre><pre>curl https://api.siray.ai/v1/images/generations/async/&lt;task_id&gt; \<br>  -H &quot;Authorization: Bearer $SIRAY_API_KEY&quot;</pre><p>Submit returns a task_id; poll every 3–5 seconds until SUCCESS, then download from data.outputs. To switch models, change only the model value.</p><h3>Tested side by side</h3><p>All tests: quality: medium, n: 1, PNG, November 20, 2026. Each prompt ran on both models at the same time.</p><ul><li><strong>White-background product</strong> — Size: 1024×1024; Flare time: 47 s; Sunburst time: 21 s; Cost each: $0.039</li><li><strong>Holiday poster with text</strong> — Size: 1024×1536; Flare time: 67 s; Sunburst time: 21 s; Cost each: $0.031</li><li><strong>Gift flat lay</strong> — Size: 1536×1024; Flare time: 46 s; Sunburst time: 139 s; Cost each: $0.031</li><li><strong>Color edit</strong> — Size: 1024×1024; Flare time: 120 s; Sunburst time: 21 s; Cost each: ≈ $0.039 + input</li></ul><h3>White-background product photo</h3><pre>E-commerce product photo of a matte black ceramic pour-over coffee kettle with a slim gooseneck spout and a natural oak handle, centered on a seamless pure white background, soft studio lighting from the upper left, a gentle contact shadow under the base, crisp edges, accurate materials, catalog style, square framing. No text, no logos, no props.</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*t41Z35J7g1E6ZKCk.jpg" /></figure><p>Both are usable catalog shots. Sunburst made the kettle larger in the frame with a slightly softer shadow; Flare left more white space. Neither added props or text.</p><h3>Holiday sale poster with exact text</h3><pre>Vertical holiday sale poster for a fictional candle shop. Three amber glass jar candles with lit flames on a dark walnut table, scattered pine sprigs and warm bokeh lights behind. At the top, large clean serif headline that reads exactly: &quot;WINTER SALE&quot;. Below it, a smaller line that reads exactly: &quot;Nov 27 – Dec 1&quot;. At the bottom, a small line that reads exactly: &quot;Northwind Candle Co.&quot;. All text spelled exactly as given, nothing else written anywhere. Warm cinematic lighting, premium print design.</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*glD8Q_B9IfTUbiHA.jpg" /></figure><p>All three lines are spelled correctly on both, including the en dash. Flare used upright serif for the date; Sunburst set it in italics and added a knit blanket in the corner, which the prompt didn’t ask for.</p><h3>Gift set flat lay</h3><pre>Overhead flat lay of a holiday gift set on a cream linen tablecloth: an open kraft gift box holding a navy wool scarf, a pair of leather gloves and a small brass thermos, wrapped with a red satin ribbon, surrounded by dried orange slices, cinnamon sticks and a few pine cones. Soft window daylight, natural shadows, realistic textures, horizontal framing. No text, no logos.</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*BflKU4KqYywf4zyn.jpg" /></figure><p>Very close. Sunburst added window-light shadows across the cloth and a herringbone weave on the scarf; Flare’s light is flatter. This was the one test where Flare finished first.</p><h3>Edit: change only the color</h3><pre>Change only the kettle body color from matte black to matte forest green. Keep the oak handle, the shape, the spout, the white background, the lighting and the shadow exactly the same. Do not add text or logos.</pre><p>Each model edited its own kettle from the first test with its -edit endpoint.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*711q6Xr3o32oAekn.jpg" /></figure><p>Both changed the body to green and kept the handle, spout and framing. Sunburst also turned the lid green; Flare kept the lid rim dark. If parts must stay untouched, send a mask.</p><h3>So which should you use?</h3><ul><li><strong>Quick drafts of many concepts</strong> — Choose: Either, test both; Why: Same price; in our runs speed varied by task more than by model</li><li><strong>Repeated edits of one image</strong> — Choose: Sunburst; Why: OpenAI positions it for consistency across edits; our single edit was clean on both</li><li><strong>Exact text on a poster</strong> — Choose: Either; Why: Both spelled all three lines correctly</li><li><strong>Lowest cost</strong> — Choose: Either at low quality; Why: Prices are identical; quality setting and size change the cost, not the model</li></ul><p>Because the price is the same, the practical approach is to send one representative prompt to both, time it on your own traffic and keep the one that wins. Switching costs nothing but a model name.</p><h3>FAQ</h3><h3>What is the difference between GPT Image 2.6 Flare and Sunburst?</h3><p>OpenAI positions Flare for fast iteration and Sunburst for higher-quality generation and more consistent edits. Both take the same parameters and support GIF and SVG output.</p><h3>Do GPT Image 2.6 Flare and Sunburst cost the same?</h3><p>Yes. On Siray AI their pricing tables are identical, for example $0.031 per 1024×1536 image and $0.039 per 1024×1024 image at medium quality (November 20, 2026).</p><h3>Is GPT Image 2.6 Flare faster than Sunburst?</h3><p>Not in our tests. Sunburst finished 3 of 4 tasks faster; Flare was faster on one. The tasks ran in parallel, so results can depend on queueing. Test on your own workload.</p><h3>Can GPT Image 2.6 render text in images?</h3><p>Yes. Both models rendered a three-line poster (“WINTER SALE”, “Nov 27 — Dec 1”, “Northwind Candle Co.”) with correct spelling in our test.</p><h3>Does GPT Image 2.6 support image editing?</h3><p>Yes, through the -edit endpoints. They accept 1 to 16 input images in images and an optional mask for inpainting.</p><h3>Try GPT Image 2.6 on Siray AI</h3><p>Both models are on one API key, billed per image with no subscription. Start from the <a href="https://console.siray.ai/model-detail/openai/gpt-image-2.6-sunburst-t2i?ref=blog.siray.ai">GPT Image 2.6 Sunburst model page</a> or compare it with Google’s model in our <a href="https://blog.siray.ai/nano-banana-pro-api-pricing-prompts/">Nano Banana Pro API guide</a>.</p><h3>Sources</h3><ul><li>OpenAI, <a href="https://x.com/OpenAI/status/2118124964197114039">announcement of GPT-Image-2.6 Flare and Sunburst on X</a>, November 2026 (seen via search; release date November 4, 2026).</li><li>Siray AI API reference: <a href="https://docs.siray.ai/api-reference/model-api/gpt-image-2.6-flare-t2i">GPT Image 2.6 Flare t2i</a>, <a href="https://docs.siray.ai/api-reference/model-api/gpt-image-2.6-flare-edit">Flare edit</a>, <a href="https://docs.siray.ai/api-reference/model-api/gpt-image-2.6-sunburst-t2i">Sunburst t2i</a>, accessed November 20, 2026.</li><li>Siray AI pricing API, GET /v1/models/pricing/{model}, November 20, 2026.</li><li>Siray AI test runs, November 20, 2026.</li></ul><p><em>Originally published on the </em><a href="https://blog.siray.ai/gpt-image-2-6-flare-vs-sunburst/"><em>Siray AI blog</em></a><em>. Prices were checked on 2026–11–20 and can change; see the live model page for current pricing.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0d139d76e3d9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GTA 6 Style AI Art: 6 Tested Prompts for the Loading-Screen Look]]></title>
            <link>https://medium.com/@siary_ai/gta-6-style-ai-art-6-tested-prompts-for-the-loading-screen-look-f920eb72ea48?source=rss-81e778058314------2</link>
            <guid isPermaLink="false">https://medium.com/p/f920eb72ea48</guid>
            <dc:creator><![CDATA[Siray AI]]></dc:creator>
            <pubDate>Fri, 20 Nov 2026 07:13:23 GMT</pubDate>
            <atom:updated>2026-11-20T07:13:23.330Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*0Gh9v3Mz--Yrdoth.png" /></figure><p>GTA 6 style AI art is the bold poster look fans know from the game’s loading screens: heavy black ink outlines, flat cel shading, halftone dots and hot-pink sunset skies over a sunny coastal city. Grand Theft Auto VI launched on November 19, 2026, and you can get that look with an image model by describing the style instead of the game. Below are six prompts we tested on three models on Siray AI, from $0.02 per image.</p><h3>Key takeaways</h3><ul><li><strong>Describe the style, not the brand.</strong> Words like “thick black ink outlines, flat cel shading, halftone dots, hot-pink and tangerine sunset” got all three models to the look. We never used the game’s name, logo or characters in a prompt.</li><li><strong>Nano Banana Pro was the fastest:</strong> 21–26 seconds per image at 1K, for $0.098.</li><li><strong>Seedream 5.5 was the cheapest:</strong> $0.02 per image (36–73 seconds), and it followed the full scene description most closely.</li><li><strong>Midjourney V8.2 was the most painterly</strong> and returns four candidates per task (118–150 seconds), listed at about $0.021.</li><li><strong>Photo-to-poster works.</strong> Both Seedream 5.5 and Nano Banana Pro turned a portrait into the poster style and kept the face and shirt, for $0.028 and $0.098.</li></ul><h3>What makes the look</h3><p>You don’t need the game’s art to describe its style. Five traits do most of the work:</p><ul><li><strong>Line work</strong> — Prompt words that worked: “thick black ink outlines”, “bold outlines”</li><li><strong>Shading</strong> — Prompt words that worked: “flat cel shading”, “flat shadows”</li><li><strong>Texture</strong> — Prompt words that worked: “subtle halftone dots”, “halftone texture”</li><li><strong>Palette</strong> — Prompt words that worked: “hot-pink and tangerine sunset sky”, “magenta and cyan neon”</li><li><strong>Setting</strong> — Prompt words that worked: “sun-bleached Florida-inspired beachfront boulevard”, “pastel art-deco hotels”, “palm silhouettes”</li></ul><p>End every prompt with “No text, no letters, no logos, no weapons.” Image models like to invent signs and lettering, and you don’t want a made-up logo on something you post.</p><h3>Models and prices on Siray AI</h3><ul><li><strong>Seedream 5.5 (ByteDance)</strong> — Endpoint: bytedance/seedream-5.5-t2i-spicy; Price per image: $0.02 at 1K sizes; Time in our tests: 36–73 s</li><li><strong>Seedream 5.5 image-to-image</strong> — Endpoint: bytedance/seedream-5.5-i2i-spicy; Price per image: $0.02 + $0.008 per input image; Time in our tests: 47 s</li><li><strong>Nano Banana Pro (Google)</strong> — Endpoint: google/nano-banana-pro-t2i / -i2i; Price per image: $0.098 at 1K or 2K $0.14; Time in our tests: 21–26 s</li><li><strong>Midjourney V8.2</strong> — Endpoint: midjourney/midjourney-v8.2-t2i; Price per image: about $0.021; Time in our tests: 118–150 s, 4 candidates</li></ul><p>Prices come from the Siray AI pricing API on November 20, 2026. Nano Banana Pro showed a 30% discount at time of writing (list price crossed out). Seedream 5.5 is listed on Siray AI only under its -spicy endpoint IDs; all our prompts were ordinary, fully clothed scenes. We ran all 17 text-to-image tasks at the same time, so timings include some queueing.</p><p>One thing to know about sizes: Seedream 5.5 returned slightly different pixel sizes than requested (864×1152 came back as 886×1182, 1424×800 as 1365×768). The aspect ratio stays close, but crop to exact sizes yourself if you need them.</p><h3>Quickstart: generate one image with the Siray API</h3><p>All three models use the same async image endpoint. Submit a task, then poll it:</p><pre>curl -X POST https://api.siray.ai/v1/images/generations/async \<br>  -H &quot;Authorization: Bearer $SIRAY_API_KEY&quot; \<br>  -H &quot;Content-Type: application/json&quot; \<br>  -d &#39;{<br>    &quot;model&quot;: &quot;bytedance/seedream-5.5-t2i-spicy&quot;,<br>    &quot;prompt&quot;: &quot;Bold graphic loading-screen poster illustration of an adult woman in oversized white sunglasses ... No text, no letters, no logos, no weapons.&quot;,<br>    &quot;size&quot;: &quot;864x1152&quot;,<br>    &quot;output_format&quot;: &quot;png&quot;<br>  }&#39;</pre><pre>curl https://api.siray.ai/v1/images/generations/async/&lt;task_id&gt; \<br>  -H &quot;Authorization: Bearer $SIRAY_API_KEY&quot;</pre><p>Poll every 3–5 seconds until the status is SUCCESS, then download the image URLs from data.outputs. For Nano Banana Pro, send &quot;size&quot;: &quot;1k&quot; and an aspect_ratio such as &quot;3:4&quot;. For Midjourney V8.2, send only aspect_ratio (it does not take size) and, optionally, a negative_prompt. Field names follow the Siray AI API reference for each endpoint.</p><h3>The 6 prompts</h3><h3>1. Loading-screen portrait</h3><pre>Bold graphic loading-screen poster illustration of an adult woman in oversized white sunglasses and a gold hoop earring, leaning on the door of a cherry-red 1980s convertible, sun-bleached Florida-inspired beachfront boulevard behind her with palm trees and pastel art-deco hotels, hot-pink and tangerine sunset sky, thick black ink outlines, flat cel shading with subtle halftone dots, high contrast, vertical composition. No text, no letters, no logos, no weapons.</pre><ul><li>Sizes: Seedream 864x1152, Nano Banana Pro 1k + 3:4, Midjourney 3:4</li><li>Time: Seedream 36 s · Nano Banana Pro 26 s · Midjourney 145 s</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YlJax0hwXxZ21Urx.jpg" /></figure><p>Seedream 5.5 and Nano Banana Pro drew the whole scene, car and hotels included. Midjourney V8.2 zoomed into a dramatic close-up on all four candidates. That is closer to a character card but loses the street. If you want the full scene from Midjourney, add “full-body, wide shot”.</p><h3>2. Four-panel collage poster</h3><pre>Four-panel collage poster separated by thin white borders, flat painted illustration style with heavy ink outlines and halftone shading: top left an airboat speeding through a mangrove swamp at sunset, top right a neon-lit nightclub entrance with a velvet rope and adult guests, bottom left a pink lifeguard tower on a crowded beach, bottom right a muscle car parked under a palm tree at dusk. Saturated pink, orange and teal palette, satirical tropical crime-drama mood without any violence. No text, no letters, no logos, no weapons.</pre><ul><li>Size: square (1024x1024 / 1:1)</li><li>Time: Seedream 36 s · Nano Banana Pro 21 s · Midjourney 134 s</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qxoOKmaOBICbq713.jpg" /></figure><p>All three models kept the four panels in the right order. This is the closest to the classic loading-screen layout. Midjourney V8.2 went for a muted, risograph-like palette; Nano Banana Pro gave the cleanest comic-book lines.</p><h3>3. Neon night street</h3><pre>Rain-slick coastal city street at night in a stylized loading-screen illustration style, magenta and cyan neon signs made of abstract shapes, a lone adult man in a loose floral shirt walking past a 24-hour diner, palm silhouettes against a purple sky, thick ink linework, flat shadows, halftone texture, cinematic wide framing. No text, no letters, no logos, no weapons.</pre><ul><li>Size: widescreen (Seedream 1424x800, others 16:9)</li><li>Time: Seedream 57 s · Nano Banana Pro 21 s · Midjourney 118 s</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*0kfmJc-pCLZZ0N78.jpg" /></figure><p>“Neon signs made of abstract shapes” kept readable words out of the signs. Midjourney V8.2 was darker and more cinematic, closer to a film still than an illustration; Seedream 5.5 kept the ink outlines strongest.</p><h3>4. Swamp town satire</h3><pre>Satirical poster illustration of a sleepy swamp town in a fictional sunny state: an alligator sunbathing beside a roadside souvenir stand covered in flamingo statues, an adult tourist in socks and sandals taking a selfie, humid golden afternoon light, bold outlines, flat cel shading, halftone dots, playful exaggerated proportions. No text, no letters, no logos, no weapons.</pre><ul><li>Size: 1152x864 / 4:3</li><li>Time: Seedream 67 s · Nano Banana Pro 26 s · Midjourney 150 s</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*6FSMv2VUiYYr0ZXi.jpg" /></figure><p>The funniest set. Nano Banana Pro painted a readable “SOUVENIRS” sign even though the prompt said no letters, so check for text before you post. “Playful exaggerated proportions” made Seedream 5.5’s tourist small and cartoonish. Drop that phrase if you want realistic adults.</p><h3>5. Turn your own photo into a poster</h3><pre>Redraw this photo as a bold graphic loading-screen poster illustration: keep the person&#39;s face, hairstyle, clothing and pose, add thick black ink outlines, flat cel shading with halftone dots, and replace the background with a hot-pink to tangerine sunset gradient with palm tree silhouettes. No text, no letters, no logos.</pre><ul><li>Input: a portrait we generated with Seedream 5.5 ($0.02), sent as a data URL in images</li><li>Endpoints: bytedance/seedream-5.5-i2i-spicy (864x1152) and google/nano-banana-pro-i2i (1k, 3:4)</li><li>Time and cost: Seedream 47 s, $0.028 · Nano Banana Pro 26 s, $0.098</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yRn01qtzn1m9lu0Q.jpg" /></figure><p>Both kept the face, curly hair and mint shirt. Seedream 5.5 replaced the whole background as asked; Nano Banana Pro kept the boardwalk and boats and restyled them, which looks more like a scene from a game. Use only photos of yourself or people who agreed to it.</p><h3>6. Phone-vlog still</h3><pre>Vertical smartphone video still, a sunburned adult man in his forties filming himself on a beach boardwalk, slightly overexposed midday light, lens flare, a seagull stealing a hot dog from a bench in the background, chaotic tropical-tourist energy, slight motion blur, realistic. Adults only. No text overlays, no letters, no logos.</pre><ul><li>Size: vertical (Seedream 800x1424, others 9:16)</li><li>Time: Seedream 73 s · Nano Banana Pro 26 s · Midjourney 134 s</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*WHM3XURqzXPGfjy5.jpg" /></figure><p>This one goes for the game’s in-world social-media clips rather than the posters. Nano Banana Pro added a phone camera interface on top of the image, which sells the idea. Midjourney V8.2 mostly dropped the seagull. This is a good first frame for an image-to-video clip.</p><h3>Prompting tips</h3><ul><li><strong>Name traits, not titles.</strong> Style words travel across models; brand names can pull in logos and copyrighted designs.</li><li><strong>Put “no text” last and check anyway.</strong> One of our 15 text-to-image results still had a readable sign.</li><li><strong>Write “adult” for every person.</strong> It keeps cartoon proportions from making people look young.</li><li><strong>Pick the model by the job:</strong> Nano Banana Pro for speed, Seedream 5.5 for the cheapest exact scene, Midjourney V8.2 when you want four painterly options to choose from.</li><li><strong>Leave out guns and crime.</strong> The look works without them, and violent images get limited reach on most social platforms.</li></ul><h3>FAQ</h3><h3>How do I make GTA 6 style AI art?</h3><p>Describe the visual traits in your prompt: thick black ink outlines, flat cel shading, halftone dots, a hot-pink and tangerine sunset, palm trees and pastel art-deco buildings. Then run it on an image model such as Seedream 5.5, Nano Banana Pro or Midjourney V8.2.</p><h3>Which AI model is best for the GTA loading-screen style?</h3><p>In our tests Seedream 5.5 followed full scenes most closely at $0.02 per image, Nano Banana Pro was fastest at 21–26 seconds, and Midjourney V8.2 gave the most painterly results with four candidates per task. Try the same prompt on two models and keep the one you like.</p><h3>Can I turn my photo into a GTA 6 style poster?</h3><p>Yes. Use an image-to-image endpoint such as bytedance/seedream-5.5-i2i-spicy or google/nano-banana-pro-i2i, send your photo in images, and ask the model to keep your face and clothing while redrawing it in the poster style. It cost us $0.028 and $0.098.</p><h3>How much does it cost to generate these images?</h3><p>On Siray AI on November 20, 2026: $0.02 per image with Seedream 5.5, $0.098 with Nano Banana Pro and about $0.021 with Midjourney V8.2. The 15 text-to-image results in this guide come to about $0.70 at those listed prices.</p><h3>Can I use the game’s logo or characters?</h3><p>Don’t. Logos, title lettering and characters belong to the game’s publisher. The prompts here describe a general art style and invent their own scenes and people.</p><h3>Try it on Siray AI</h3><p>One API key gives you Seedream 5.5, Nano Banana Pro and Midjourney V8.2, billed per image with no subscription. Start from the <a href="https://console.siray.ai/model-detail/bytedance/seedream-5.5-t2i-spicy?ref=blog.siray.ai">Seedream 5.5 model page</a>, or read our <a href="https://blog.siray.ai/seedream-5-5-api-pricing-prompts/">Seedream 5.5 API guide</a> and <a href="https://blog.siray.ai/nano-banana-pro-api-pricing-prompts/">Nano Banana Pro API guide</a>.</p><p><em>Not affiliated with or endorsed by Rockstar Games or Take-Two Interactive. Grand Theft Auto is a trademark of Take-Two Interactive. All images in this article were generated from the text prompts shown.</em></p><h3>Sources</h3><ul><li>Xbox Wire, <a href="https://news.xbox.com/en-us/2026/11/10/grand-theft-auto-6-comes-to-xbox-cloud-gaming-november-19/">Grand Theft Auto VI comes to Xbox Cloud Gaming on November 19</a>, published November 10, 2026.</li><li>Forbes, <a href="https://www.forbes.com/sites/siladityaray/2026/11/18/grand-theft-auto-6-launches-to-rave-reviews-over-15-billion-in-estimated-pre-sales/">Grand Theft Auto 6 launches to rave reviews</a>, November 18, 2026.</li><li>Siray AI API reference: <a href="https://docs.siray.ai/api-reference/model-api/seedream-5.5-t2i-spicy">Seedream 5.5 t2i Spicy</a>, <a href="https://docs.siray.ai/api-reference/model-api/seedream-5.5-i2i-spicy">Seedream 5.5 i2i Spicy</a>, <a href="https://docs.siray.ai/api-reference/model-api/nano-banana-pro-t2i">Nano Banana Pro t2i</a>, <a href="https://docs.siray.ai/api-reference/model-api/nano-banana-pro-i2i">Nano Banana Pro i2i</a>, <a href="https://docs.siray.ai/api-reference/model-api/midjourney-v8.2-t2i">Midjourney V8.2 t2i</a>, accessed November 20, 2026.</li><li>Siray AI pricing API, GET /v1/models/pricing/{model}, November 20, 2026.</li><li>Siray AI test runs, November 20, 2026.</li></ul><p><em>Originally published on the </em><a href="https://blog.siray.ai/gta-6-style-ai-art-prompts/"><em>Siray AI blog</em></a><em>. Prices were checked on 2026–11–20 and can change; see the live model page for current pricing.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f920eb72ea48" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Is Image-to-Video? How i2v APIs Work (Inputs, Pricing, Polling)]]></title>
            <link>https://medium.com/@siary_ai/what-is-image-to-video-how-i2v-apis-work-inputs-pricing-polling-e8b681c07845?source=rss-81e778058314------2</link>
            <guid isPermaLink="false">https://medium.com/p/e8b681c07845</guid>
            <category><![CDATA[apiai]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Siray AI]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 08:25:11 GMT</pubDate>
            <atom:updated>2026-11-19T08:25:11.963Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lrVPDqfgo5g99npa.png" /></figure><p>Image-to-video (i2v) is a type of AI video generation that turns one still image into a short clip. You send the image and a text prompt describing the motion, and the model animates that exact picture. On Siray AI, a 5-second 720p i2v clip costs $0.47 to $1.305, depending on the model.</p><h3>Key takeaways</h3><ul><li><strong>The image is the first frame.</strong> The clip starts on your picture, so the face, product and lighting stay the same while the prompt adds motion.</li><li><strong>Some models also take a last frame.</strong> Kling 4.0 and Seedance 2.5 accept an optional end_image, so you control where the clip ends. Kling 4.0 Flash does not.</li><li><strong>Billing is per second of output.</strong> 5 seconds at 720p costs $0.47 on Kling 4.0 Flash, $0.785 on Kling 4.0 and $1.305 on Seedance 2.5 (November 19, 2026).</li><li><strong>The output takes the shape of your image.</strong> Our portrait rooftop photo gave an 836×1116 portrait clip. Seedance 2.5 rejected an explicit 9:16 and needed adaptive.</li><li><strong>i2v APIs are asynchronous.</strong> You submit a task, get a task_id and poll until it finishes, which took about 3 to 5 minutes in our tests.</li></ul><h3>Image-to-video vs text-to-video vs reference-to-video</h3><ul><li><strong>Text-to-video (t2v)</strong> — What you send: A prompt; What stays fixed: Nothing, the model invents the scene; Typical use: Concept clips, b-roll, ideas</li><li><strong>Image-to-video (i2v)</strong> — What you send: One image (plus an optional last frame) and a prompt; What stays fixed: The first frame, exactly; Typical use: Animating a product photo, portrait or artwork</li><li><strong>Reference-to-video (ref2v)</strong> — What you send: Several reference images or videos and a prompt; What stays fixed: The look of people or objects, not a fixed frame; Typical use: Putting the same character or product into new scenes</li></ul><p>Choose i2v when you already have the exact picture you want to start from. Choose ref2v when you want the same subject in a new composition. Choose t2v when you have no image at all.</p><h3>What you send to an image-to-video API</h3><p>Most i2v endpoints on Siray AI share the same request shape. These fields come from the Siray AI API reference for each endpoint:</p><ul><li><strong>image</strong> — What it does: First frame, as a URL or data URL; Kling 4.0 i2v: Required; Kling 4.0 Flash i2v: Required; Seedance 2.5 i2v: Required</li><li><strong>prompt</strong> — What it does: Describes the motion and camera; Kling 4.0 i2v: Required; Kling 4.0 Flash i2v: Required; Seedance 2.5 i2v: Required</li><li><strong>end_image</strong> — What it does: Optional last frame; Kling 4.0 i2v: Yes; Kling 4.0 Flash i2v: No; Seedance 2.5 i2v: Yes</li><li><strong>duration</strong> — What it does: Clip length in seconds; Kling 4.0 i2v: 3–30; Kling 4.0 Flash i2v: 3–20; Seedance 2.5 i2v: 4–30</li><li><strong>size</strong> — What it does: Output resolution; Kling 4.0 i2v: 720p, 1080p; Kling 4.0 Flash i2v: 720p; Seedance 2.5 i2v: 480p, 720p, 1080p</li><li><strong>aspect_ratio</strong> — What it does: Output shape; Kling 4.0 i2v: auto or a fixed ratio; Kling 4.0 Flash i2v: auto or a fixed ratio; Seedance 2.5 i2v: adaptive (see below)</li><li><strong>Audio</strong> — What it does: Sound in the clip; Kling 4.0 i2v: Always on in our tests; Seedance 2.5 i2v: audio_enable true or false</li></ul><p><strong>Write the prompt about motion, not appearance.</strong> The image already shows what things look like. Describe what moves, how the camera moves, and what must stay the same.</p><h3>A tested example: first and last frame</h3><p>We sent Kling 4.0’s i2v endpoint two images of the same woman on a rooftop. The first shows her in a green wool coat. The last frame, which we edited from the first with Seedream 5.5, shows her without the coat, which is folded on the wall.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/836/0*DPw_uijrSt24-F43.jpg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7fpborqzsFcRU7ot.jpg" /></figure><pre>The woman on the rooftop slowly unbuttons her moss-green wool coat, slips it off her shoulders, folds it once and lays it on the white parapet wall to her right, leaving her in the cream turtleneck sweater and charcoal trousers. She then turns back toward the camera and settles into the same pose as the end frame. Warm golden-hour light, gentle breeze in her hair, steady handheld camera. Adult woman only, fully clothed. No text, no logos.</pre><ul><li>Endpoint: kling-4.0-i2v with end_image, 5 s, 720p, aspect_ratio: auto</li><li>Time: 284 seconds · Cost: $0.785 · Output: 836×1116, the same shape as the input images</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/836/0*vT700r_wfUXcfxG5.jpg" /></figure><p><a href="https://blog.siray.ai/content/media/2026/11/kling-4-0-i2v-coat-off.mp4">Watch the clip</a></p><p>The clip opens on the first image and ends on the last one, with the same face, sweater, rooftop and light. In between, the model invents the action. For about a second the coat stands stiffly on the wall before it settles. <strong>The last frame should differ from the first in shape, not only in color.</strong> An earlier try with a same-cut blazer in another color looked like a color swap, not a coat coming off. The full test is in our <a href="https://blog.siray.ai/kling-4-0-api-pricing-prompts/">Kling 4.0 API guide</a>.</p><h3>Why the output takes the shape of your image</h3><p>An i2v clip starts on your picture, so most models keep its proportions:</p><ul><li><strong>Kling 4.0</strong> with aspect_ratio: auto returned an 836×1116 clip, matching the portrait shape of the input photos.</li><li><strong>Seedance 2.5</strong> rejected aspect_ratio: &quot;9:16&quot; on a 9:16 photo with the error &quot;For first-frame or first-last-frame generation, the output ratio follows the first-frame image.&quot; With adaptive it returned a 720×1280 clip. The API reference still lists 9:16 as a valid value, so use adaptive for Seedance 2.5 i2v.</li></ul><p><strong>Crop or generate the image in the ratio you need before you call the API.</strong> A 9:16 photo gives a vertical clip for Reels and TikTok; a 16:9 photo gives a widescreen one. We used this with Seedance 2.5 in <a href="https://blog.siray.ai/black-friday-product-video-ads-seedance-2-5/">Black Friday Video Ads from One Product Photo</a>.</p><h3>How image-to-video pricing works</h3><p>Siray AI bills i2v by the seconds of video you get back. The input image doesn’t add length. Prices from the Siray AI pricing API on November 19, 2026:</p><ul><li><strong>Kling 4.0 Flash i2v</strong> — 5 s at 720p: $0.47; 10 s at 720p: $0.94; Per second at 720p: ≈ $0.094</li><li><strong>Kling 4.0 i2v</strong> — 5 s at 720p: $0.785; 10 s at 720p: $1.57; Per second at 720p: ≈ $0.157</li><li><strong>Seedance 2.5 i2v</strong> — 5 s at 720p: $1.305 $1.45; 10 s at 720p: $2.61 $2.90; Per second at 720p: ≈ $0.261</li></ul><p>Seedance 2.5 showed a 10% discount at time of writing (list price crossed out). Its price is calculated from tokens: height × width × (input duration + output duration) × 24 / 1024, at $0.012038 per 1K tokens. A still image has no duration, so only the output counts. A 720×1280 clip of 5 seconds is 108,000 tokens, about $1.30.</p><p>Two practical rules:</p><ul><li><strong>Draft cheap, finish expensive.</strong> Test motion at 480p or on a Flash tier, then re-run the take you like at full quality.</li><li><strong>Length costs more than you expect.</strong> 10 seconds costs twice as much as 5. Many ads only need 5.</li></ul><p>Check the live prices on the <a href="https://console.siray.ai/?ref=blog.siray.ai">Siray AI console</a> before a batch.</p><h3>How the async API flow works</h3><p>Video generation takes minutes, so i2v APIs don’t make you wait on one HTTP request. They work in two steps.</p><p><strong>1. Submit the task</strong></p><pre>curl -X POST https://api.siray.ai/v1/video/generations \<br>  -H &quot;Authorization: Bearer $SIRAY_API_KEY&quot; \<br>  -H &quot;Content-Type: application/json&quot; \<br>  -d &#39;{<br>    &quot;model&quot;: &quot;kuaishou/kling-4.0-i2v&quot;,<br>    &quot;image&quot;: &quot;https://your-cdn.example.com/rooftop-coat.jpg&quot;,<br>    &quot;end_image&quot;: &quot;https://your-cdn.example.com/rooftop-no-coat.jpg&quot;,<br>    &quot;prompt&quot;: &quot;The woman on the rooftop slowly unbuttons her moss-green wool coat...&quot;,<br>    &quot;duration&quot;: 5,<br>    &quot;size&quot;: &quot;720p&quot;,<br>    &quot;aspect_ratio&quot;: &quot;auto&quot;<br>  }&#39;</pre><p>The response contains a task_id. In our tests it came back at the top level of the response, not under data as the API reference shows, so read both.</p><p><strong>2. Poll until it finishes</strong></p><pre>curl https://api.siray.ai/v1/video/generations/&lt;task_id&gt; \<br>  -H &quot;Authorization: Bearer $SIRAY_API_KEY&quot;</pre><p>Check every 3 to 5 seconds. data.status moves through NOT_START, SUBMITTED, QUEUED and IN_PROGRESS to SUCCESS or FAILURE. On success, data.outputs</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e8b681c07845" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Runway API Alternatives for Developers: Kling 4.0, Seedance 2.5, Veo 3.1 and More on One API]]></title>
            <link>https://medium.com/@siary_ai/runway-api-alternatives-for-developers-kling-4-0-seedance-2-5-veo-3-1-and-more-on-one-api-d53aa88c17fa?source=rss-81e778058314------2</link>
            <guid isPermaLink="false">https://medium.com/p/d53aa88c17fa</guid>
            <category><![CDATA[runway]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Siray AI]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 08:16:13 GMT</pubDate>
            <atom:updated>2026-11-19T08:16:13.545Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Jq6l7oIvAHRmTqP-.png" /></figure><p>A Runway API alternative only makes sense model by model. Runway’s API now resells Kling 4.0, Seedance 2.5, Wan 3 and Veo 3.1, and Siray AI offers the same models through one key. On November 18, 2026, Seedance 2.5 and Wan 3.0 cost less per second on Siray AI, while Kling 4.0 cost less on Runway during its promotion.</p><h3>Key takeaways</h3><ul><li><strong>Same models, different prices:</strong> a 5-second 720p Seedance 2.5 clip costs $1.50 on Runway and $1.305 on Siray AI. A 5-second 720p Wan 3.0 clip costs $0.50 on Runway and $0.40 on Siray AI.</li><li><strong>Kling 4.0 is cheaper on Runway until November 30:</strong> Runway’s 20% promotion brings Kling 4.0 at 720p to $0.1256 per second, against $0.157 on Siray AI. At list price the two match.</li><li><strong>Runway’s own models have no equivalent:</strong> Gen-4.5, Aleph 2.0 and Act-Two are only on Runway. If your pipeline depends on them, stay there.</li><li><strong>One prompt, four models:</strong> we ran the same 5-second 720p camera-move prompt on Kling 4.0, Kling 4.0 Flash, Seedance 2.5 and Wan 3.0 through Siray AI. Jobs took 87 to 223 seconds and cost $0.40 to $1.305.</li><li><strong>Discounts move both ways:</strong> several Siray AI prices below include a discount at the time of writing. At list price, Wan 3.0 on Siray AI would cost more than on Runway.</li></ul><h3>What Runway’s API sells today</h3><p>Runway’s <a href="https://docs.dev.runwayml.com/guides/pricing/">API pricing page</a> bills in credits at $0.01 each, bought in the developer portal. The same page lists Runway’s own models (Gen-4.5, Gen-4 Turbo, Aleph 2.0, Act-Two) next to third-party video models: Kling 2.5 Turbo to 4.0, Seedance 2.0 and 2.5, WAN 3, Veo 3.1 and 3.1 Fast, MiniMax Hailuo, Grok Imagine and others.</p><p>That changes the question. For most of these models, choosing a Runway alternative is no longer about the model. It is about the price, rounding and minimum charges, the endpoints available for that model, and whether you also need Runway’s own tools.</p><h3>Runway API alternative pricing: the same models per second</h3><p>All prices are for output video without a video input, in US dollars per second. Runway prices are converted from credits at $0.01 per credit. Both were checked on November 18, 2026.</p><ul><li><strong>Kling 4.0</strong> — Resolution: 720p; Runway API: $0.157 list, $0.1256 with promotion; Siray AI: $0.157</li><li><strong>Kling 4.0</strong> — Resolution: 1080p; Runway API: $0.235 list, $0.188 with promotion; Siray AI: $0.235</li><li><strong>Kling 4.0 Flash</strong> — Resolution: 720p; Runway API: $0.094; Siray AI: $0.094</li><li><strong>Seedance 2.5</strong> — Resolution: 480p; Runway API: $0.20; Siray AI: ≈ $0.121</li><li><strong>Seedance 2.5</strong> — Resolution: 720p; Runway API: $0.30; Siray AI: ≈ $0.261</li><li><strong>Seedance 2.5</strong> — Resolution: 1080p; Runway API: $0.68; Siray AI: ≈ $0.640</li><li><strong>Seedance 2.5</strong> — Resolution: 4K; Runway API: $1.10; Siray AI: not offered</li><li><strong>Wan 3.0 (Runway: </strong><strong>wan3)</strong> — Resolution: 480p / 720p / 1080p; Runway API: $0.05 / $0.10 / $0.20; Siray AI: $0.04 / $0.08 / $0.16</li><li><strong>Veo 3.1</strong> — Runway API: $0.40 with audio, $0.20 without; Siray AI: $2.56 per fixed 8-second clip</li></ul><p>Things the table does not show:</p><ul><li><strong>Promotions and discounts.</strong> Runway’s Kling 4.0 rate is 20% off “through November 30, 2026 UTC”, and list prices apply from December 1. On Siray AI, Seedance 2.5 shows 10% off (list ≈ $0.29 per second at 720p), Wan 3.0 30% off (list ≈ $0.114 per second at 720p) and Veo 3.1 20% off (list $3.20 per clip) at the time of writing. At list price, Wan 3.0 on Siray AI would cost more than Runway’s $0.10 per second.</li><li><strong>Rounding and minimums.</strong> Runway rounds Kling charges up to the next credit, has a 29-credit minimum for Kling 4.0 Flash and an 80-credit ($0.80) minimum per Seedance 2.5 generation. Siray AI bills by duration: Seedance 2.5 starts at 4 seconds, Kling 4.0 at 3 and Wan 3.0 at 2.</li><li><strong>Veo 3.1 is not directly comparable.</strong> Siray AI’s Veo 3.1 endpoint takes only prompt and a fixed duration of 8, and its API reference does not state the output resolution or whether audio is included. An 8-second clip costs $3.20 with audio or $1.60 without on Runway, and $2.56 on Siray AI. Check what you need before you compare.</li><li><strong>Video inputs cost extra on both.</strong> Runway and Siray AI both add a charge for input or reference video on Kling 4.0 and Seedance 2.5, with different formulas. See the <a href="https://blog.siray.ai/kling-4-0-api-pricing-prompts/">Kling 4.0 guide</a> for Siray AI’s.</li></ul><p><strong>Worked example: 20 five-second 720p clips, at November 18 prices</strong></p><ul><li>Kling 4.0: $12.80 on Runway (64 credits each with the promotion), $15.70 on Siray AI.</li><li>Kling 4.0 Flash: $9.40 on both.</li><li>Seedance 2.5: $30.00 on Runway, $26.10 on Siray AI.</li><li>Wan 3.0: $10.00 on Runway, $8.00 on Siray AI.</li></ul><p>For live numbers, check Runway’s pricing page and the <a href="https://console.siray.ai/?ref=blog.siray.ai">Siray AI model pages</a>.</p><p><a href="https://blog.siray.ai/content/media/2026/11/runway-alternatives-concept-per-second-billing.mp4">Watch the clip: four hourglasses of different sizes running at different speeds, a short concept video about per-second billing, made with Wan 3.0</a></p><p><em>Concept video generated with Wan 3.0 on Siray AI.</em></p><h3>Model mapping: Runway API to Siray AI</h3><ul><li><strong>kling_4</strong> — On Siray AI: kuaishou/kling-4.0-t2v, -i2v, -ref2v, -video-edit; Notes: 3–30 s, 720p or 1080p, audio always on</li><li><strong>kling_4_flash</strong> — On Siray AI: kuaishou/kling-4.0-flash-t2v, -i2v, -ref2v, -video-edit; Notes: 3–20 s, 720p</li><li><strong>Seedance 2.5</strong> — On Siray AI: bytedance/seedance-2.5-t2v, -i2v, -ref2v; Notes: 4–30 s, 480p to 1080p, audio_enable. Siray AI also has a Seedance 2.5 Flash tier</li><li><strong>wan3</strong> — On Siray AI: alibaba/wan-3.0-t2v, -i2v, -ref2v; Notes: 2–30 s, 480p to 1080p, audio_enable</li><li><strong>wan3_prime</strong> — On Siray AI: Prime is listed only as spicy endpoints</li><li><strong>veo3.1</strong> — On Siray AI: google/veo-3.1-t2v, google/veo-3.1-i2v; Notes: fixed 8 s</li><li><strong>veo3.1_fast</strong> — On Siray AI: not listed</li><li><strong>gen4.5, </strong><strong>gen4_turbo, </strong><strong>aleph2, </strong><strong>act_two</strong> — On Siray AI: not available; Notes: Runway&#39;s own models</li></ul><p>Durations, resolutions and parameter names come from the Siray AI API reference. Request formats differ between the two platforms, so moving a pipeline means changing the base URL, the key, the model IDs and the request body. A Runway SDK call will not work against Siray AI unchanged.</p><h3>One prompt on four models</h3><p>We sent the same prompt to four models through the Siray AI API on November 18, 2026, all at 5 seconds, 720p and 16:9, so the results compare like for like. Costs are calculated from Siray AI’s prices for each request at the time of the test.</p><pre>A slow, low tracking shot gliding forward along a stone path between tall palms and broad banana leaves inside a Victorian glass greenhouse at dawn. Mist hangs in the air, water drops fall from the leaves, and pale sunlight comes through the fogged glass panes. In the final second the camera rises smoothly past the white iron ribs of the glass dome to reveal the morning sky. Cinematic 35mm film look, soft grain, natural motion. No people. No text, no signs, no letters, no logos.</pre><ul><li><strong>Kling 4.0</strong> — Settings: 5 s, 720p; Time: 223 s; Cost: $0.785; Output: 1280×720, 24 fps, audio</li><li><strong>Kling 4.0 Flash</strong> — Settings: 5 s, 720p; Time: 87 s; Cost: $0.47; Output: 1280×720, 24 fps, audio</li><li><strong>Seedance 2.5</strong> — Settings: 5 s, 720p, audio_enable: true; Time: 208 s; Cost: $1.305; Output: 1280×720, 24 fps, audio</li><li><strong>Wan 3.0</strong> — Settings: 5 s, 720p, audio_enable: true; Time: 178 s; Cost: $0.40; Output: 1280×720, 30 fps, audio</li></ul><h3>Kling 4.0</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*GS9fHL_oRobtRm-e.jpg" /></figure><p><a href="https://blog.siray.ai/content/media/2026/11/kling-4-0-greenhouse-720p.mp4">Watch the clip</a></p><p>Kling 4.0 kept closest to the “stone path between tall palms” and gave the heaviest mist, in a long arched glass hall. The rise happens at the very end and passes through the palm fronds, so the last second is soft with motion blur before the ribbed vault comes into view.</p><h3>Kling 4.0 Flash</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yh7oO8zMxYcctAtJ.jpg" /></figure><p><a href="https://blog.siray.ai/content/media/2026/11/kling-4-0-flash-greenhouse-720p.mp4">Watch the clip</a></p><p>Flash finished in 87 seconds, less than half the time of Kling 4.0, and still made the glide and the rise. It built a pitched glass roof rather than a dome, with a wet flagstone floor and a pink dawn sky at the end of the house.</p><h3>Seedance 2.5</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*aITweSuAgMo9IiFv.jpg" /></figure><p><a href="https://blog.siray.ai/content/media/2026/11/seedance-2-5-greenhouse-720p.mp4">Watch the clip</a></p><p>Seedance 2.5 was the only model to end on a true curved dome, with scrolled ironwork and sunlight breaking through the mist, the closest match to “Victorian glass dome”. The trade-off is the path: the plants and mist are so dense that the stone path is only visible for the first moment.</p><h3>Wan 3.0</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*d6tWEa7WgHLiStWT.jpg" /></figure><p><a href="https://blog.siray.ai/content/media/2026/11/wan-3-0-greenhouse-720p.mp4">Watch the clip</a></p><p>Wan 3.0 gave the brightest and cleanest frame, with a wet path, decorative iron arches and a clear tilt up to the morning sky through the glass. Like Flash it chose a pitched roof over a dome, and it was the cheapest job, at $0.40.</p><h3>Quickstart: submit and poll on Siray AI</h3><p>Every video model on Siray AI uses the same two calls. Submit a job to POST /v1/video/generations to get a task_id, then poll GET /v1/video/generations/{task_id} every few seconds until data.status is SUCCESS or FAILURE. The video URL is in data.outputs.</p><p><a href="https://blog.siray.ai/content/media/2026/11/runway-alternatives-concept-submit-poll.mp4">Watch the clip: film goes into a developing tank in a red-lit darkroom and comes back out developed, a short concept video about the submit-and-poll flow, made with Wan 3.0</a></p><p><em>Concept video generated with Wan 3.0 on Siray AI.</em></p><p>This is the Node.js script we used for all four runs above (Node.js 18 or later, no dependencies):</p><pre>// node generate.mjs request.json out.mp4<br>import { readFile, writeFile } from &quot;node:fs/promises&quot;;</pre><pre>const API = &quot;https://api.siray.ai/v1/video/generations&quot;;<br>const headers = {<br>  Authorization: `Bearer ${process.env.SIRAY_API_KEY}`,<br>  &quot;Content-Type&quot;: &quot;application/json&quot;,<br>};<br>const sleep = (ms) =&gt; new Promise((r) =&gt; setTimeout(r, ms));</pre><pre>const body = JSON.parse(await readFile(process.argv[2], &quot;utf8&quot;));<br>const out = process.argv[3] ?? &quot;output.mp4&quot;;</pre><pre>const submit = await (await fetch(API, { method: &quot;POST&quot;, headers, body: JSON.stringify(body) })).json();<br>const taskId = submit.task_id ?? submit.data?.task_id;<br>if (!taskId) throw new Error(`No task_id: ${JSON.stringify(submit)}`);</pre><pre>const started = Date.now();<br>while (Date.now() - started &lt; 15 * 60 * 1000) {<br>  await sleep(5000);<br>  const { data } = await (await fetch(`${API}/${taskId}`, { headers })).json();<br>  if (data.status === &quot;SUCCESS&quot;) {<br>    const url = [].concat(data.outputs)[0];<br>    await writeFile(out, Buffer.from(await (await fetch(url)).arrayBuffer()));<br>    console.log(`done in ${Math.round((Date.now() - started) / 1000)}s`);<br>    process.exit(0);<br>  }<br>  if (data.status === &quot;FAILURE&quot;) throw new Error(data.fail_reason);<br>}<br>throw new Error(`timed out; keep polling task ${taskId}`);</pre><p>And the request.json for the Wan 3.0 run:</p><pre>{<br>  &quot;model&quot;: &quot;alibaba/wan-3.0-t2v&quot;,<br>  &quot;prompt&quot;: &quot;A slow, low tracking shot gliding forward along a stone path between tall palms...&quot;,<br>  &quot;duration&quot;: 5,<br>  &quot;size&quot;: &quot;720p&quot;,<br>  &quot;aspect_ratio&quot;: &quot;16:9&quot;,<br>  &quot;audio_enable&quot;: true<br>}</pre><p>To switch models, change model and the fields that model&#39;s API reference lists. Kling 4.0 has no audio_enable, and Veo 3.1 takes only prompt and duration. Three things we noticed in testing:</p><ul><li>The task_id came back at the top level of the submit response, while the API reference shows it under data. The script reads both.</li><li>fail_reason can contain the output URL on successful tasks, so judge success by status only.</li><li>If the script times out, keep polling the same task_id instead of submitting again, or you pay twice.</li></ul><h3>When to stay on Runway, and when to switch</h3><p><strong>Stay on Runway if:</strong></p><ul><li>You use Gen-4.5, Aleph 2.0 or Act-Two. They are Runway’s own models and are not on Siray AI.</li><li>You need Seedance 2.5 at 4K, Veo 3.1 Fast or Wan 3 Prime without the spicy variants.</li><li>You generate mostly Kling 4.0 before November 30 and the promotion covers your volume.</li><li>You need professional formats such as ProRes or HDR output, which Runway offers on Gen-4.5 and Aleph 2.0.</li></ul><p><strong>Consider Siray AI if:</strong></p><ul><li>Most of your volume is Seedance 2.5 or Wan 3.0 at 480p to 1080p, where the per-second price was lower on November 18.</li><li>You want one key for these video models plus image models such as Seedream and Midjourney.</li><li>You need endpoints Runway does not list, such as Seedance 2.5 Flash or the spicy variants for adult-audience projects.</li><li>You make many short Seedance clips, where Runway’s 80-credit minimum per generation adds up.</li></ul><p>For more on each model, see our guides to <a href="https://blog.siray.ai/kling-4-0-api-pricing-prompts/">Kling 4.0</a>, <a href="https://blog.siray.ai/wan-3-0-explained-alibabas-30-second-ai-video-model/">Wan 3.0</a> and <a href="https://blog.siray.ai/google-veo-3-1/">Veo 3.1</a>. Our earlier <a href="https://blog.siray.ai/runway-alternative-comparing-ai-video-apis-price-speed/">Seedance 2.0 price comparison with Runway</a> covers the previous generation.</p><h3>FAQ</h3><h3>Is there a cheaper alternative to the Runway API?</h3><p>It depends on the model. On November 18, 2026, Seedance 2.5 and Wan 3.0 cost less per second on Siray AI, Kling 4.0 cost less on Runway during its promotion, and Kling 4.0 Flash cost the same on both.</p><h3>How much does the Runway API cost per second?</h3><p>Runway bills credits at $0.01 each. For example, Kling 4.0 Flash is 9.4 credits per second ($0.094), Seedance 2.5 at 720p is 30 credits per second ($0.30), and Veo 3.1 is 40 credits per second with audio ($0.40).</p><h3>Can I use Runway Gen-4.5 through Siray AI?</h3><p>No. Gen-4.5, Gen-4 Turbo, Aleph 2.0 and Act-Two are Runway’s own models and are only available from Runway.</p><h3>Can I switch from the Runway API to Siray AI without changing my code?</h3><p>No. The two APIs use different base URLs, keys, model IDs and request bodies. The submit-and-poll pattern is similar, so the change is usually small. The script above shows the Siray AI side.</p><h3>Which models are available on both Runway and Siray AI?</h3><p>Kling 4.0, Kling 4.0 Flash, Seedance 2.5, Wan 3.0 and Veo 3.1. The endpoints, durations and resolutions differ, so check the mapping table above.</p><h3>Try these models on Siray AI</h3><p>Create an API key in the <a href="https://console.siray.ai/?ref=blog.siray.ai">Siray AI console</a> and run the script above with any of the model IDs in the mapping table. One key covers Kling 4.0, Seedance 2.5, Wan 3.0, Veo 3.1 and the other models on Siray AI.</p><h3>Sources</h3><ul><li>Runway, <a href="https://docs.dev.runwayml.com/guides/pricing/">API pricing</a>, no update date shown, accessed November 18, 2026.</li><li>Siray AI API reference: <a href="https://docs.siray.ai/api-reference/model-api/seedance-2.5-t2v">Seedance 2.5 t2v</a>, <a href="https://docs.siray.ai/api-reference/model-api/wan-3.0-t2v">Wan 3.0 t2v</a>, <a href="https://docs.siray.ai/api-reference/model-api/veo-3.1-t2v">Veo 3.1 t2v</a>, <a href="https://docs.siray.ai/api-reference/model-api/kling-4.0-t2v">Kling 4.0 t2v</a>, accessed November 17–18, 2026.</li><li>Siray AI pricing API, GET /v1/models/pricing/{model}, November 18, 2026.</li><li>Siray AI model list, GET /v1/models, November 18, 2026.</li></ul><p><em>Originally published on the </em><a href="https://blog.siray.ai/runway-api-alternatives/"><em>Siray AI blog</em></a><em>. Prices were checked on 2026–11–18 and can change; see the live model page for current pricing.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d53aa88c17fa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Black Friday Video Ads from One Product Photo: A Seedance 2.5 Image-to-Video Workflow]]></title>
            <link>https://medium.com/@siary_ai/black-friday-video-ads-from-one-product-photo-a-seedance-2-5-image-to-video-workflow-bc98d9e50e32?source=rss-81e778058314------2</link>
            <guid isPermaLink="false">https://medium.com/p/bc98d9e50e32</guid>
            <category><![CDATA[seedance-2]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Siray AI]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 07:32:51 GMT</pubDate>
            <atom:updated>2026-11-19T07:32:51.229Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VW_MBJiF_624Kj-4.png" /></figure><p>Black Friday video ads don’t need a video shoot. With ByteDance’s Seedance 2.5 image-to-video model on Siray AI, one product photo becomes a set of 5-second vertical clips for Reels, TikTok and Shorts. In our test, each 720p clip cost <strong>$1.305</strong> and took about three minutes, and three different ad shots came from the same photo.</p><h3>Key takeaways</h3><ul><li><strong>One photo, three ads:</strong> a push-in, a camera orbit and a hands-on unboxing, all from the same 9:16 product photo.</li><li><strong>Cost:</strong> $0.035 for the product photo plus $1.305 per 5-second 720p clip, so the full set of three ads cost $3.95.</li><li><strong>Speed:</strong> each clip finished in 177–181 seconds, and all three ran at the same time.</li><li><strong>Set </strong><strong>aspect_ratio to </strong><strong>adaptive:</strong> Seedance 2.5 i2v rejected 9:16 and takes the shape of your photo instead. Make the photo vertical and the video follows.</li><li><strong>Silent by default:</strong> with audio_enable: false the MP4 has no audio track at all, which suits ads where you add licensed music in the editor.</li></ul><h3>The workflow at a glance</h3><ol><li><strong>Get one clean product photo</strong> in 9:16. Use a real packshot, or generate one (we used Seedream 5.5).</li><li><strong>Write one prompt per shot.</strong> Describe a single camera move and say what must not change.</li><li><strong>Submit all shots in parallel</strong> to bytedance/seedance-2.5-i2v with the same photo as image.</li><li><strong>Poll, download and edit:</strong> add your offer text, music and logo in a video editor, not in the prompt.</li></ol><p>Seedance 2.5 is ByteDance’s model, available on Siray AI through its unified API. ByteDance Seed describes it as “a next-generation audio-video joint generation model, built for 30-second storytelling”, with smoother motion and more realistic visuals than earlier versions.</p><h3>Step 1: the product photo</h3><p>We generated a fictional product so the ad has no real brand in it: an amber glass perfume bottle with a gold cap in an open black gift box, with holiday lights behind it.</p><pre>Vertical product photograph for a holiday ad. A rectangular amber glass perfume bottle with a brushed gold cap stands upright inside an open matte black gift box lined with crumpled ivory tissue paper. The box lid leans against the back of the box, and a loose satin ribbon in deep burgundy drapes over the front edge. The box sits on a warm white marble counter. Soft key light from the upper left, a gentle rim light behind the bottle so the amber glass glows, shallow depth of field, out-of-focus warm string lights in the background as round bokeh. The bottle and box are fully in frame with clear space above the bottle. Clean commercial studio look, photorealistic. No people, no hands. The bottle and box have no label and no printing. No text, no letters, no numbers, no logos, no watermark.</pre><ul><li>Endpoint: bytedance/seedream-5.5-t2i-spicy, size: 1584x2816, output_format: jpg</li><li>Time: 31 seconds · Cost: $0.035 · Output: 1536×2730 (slightly smaller than the requested size, still 9:16)</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*TsFyY-QRAC1aFmmy.jpg" /></figure><p>Seedream 5.5 is listed on Siray AI only as Spicy endpoints, but the prompt and the result are a plain product shot. If you already have a packshot, skip this step and use your own photo. <strong>Leave room above and around the product.</strong> Every camera move in step 2 needs space to move into.</p><h3>Step 2: three ad shots from the same photo</h3><p>We ran all three prompts on November 19, 2026 with the same request settings. Times are from submit to finish. Costs are Siray AI’s price for a 5-second 720p clip.</p><pre>{<br>  &quot;model&quot;: &quot;bytedance/seedance-2.5-i2v&quot;,<br>  &quot;image&quot;: &quot;&lt;URL or data URL of the product photo&gt;&quot;,<br>  &quot;prompt&quot;: &quot;&lt;one of the prompts below&gt;&quot;,<br>  &quot;duration&quot;: 5,<br>  &quot;size&quot;: &quot;720p&quot;,<br>  &quot;aspect_ratio&quot;: &quot;adaptive&quot;,<br>  &quot;audio_enable&quot;: false<br>}</pre><h3>Shot 1: slow push-in (the hero shot)</h3><pre>Slow, smooth dolly push-in toward the amber glass perfume bottle in the open black gift box. The camera moves straight forward from a medium shot to a close-up of the bottle and its brushed gold cap. A soft highlight glides across the gold cap and the amber glass glows from the rim light behind it. The warm string lights in the background twinkle gently as out-of-focus bokeh. The bottle, the box, the tissue paper and the burgundy ribbon stay still and keep their exact shape. Premium commercial look, steady camera, no shake. No people, no text appears.</pre><ul><li>Endpoint: seedance-2.5-i2v, 5 s, 720p, aspect_ratio: adaptive</li><li>Time: 181 seconds · Cost: $1.305 · Output: 720×1280, 24 fps, no audio track</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/0*8gvnxE--0ItnwWoo.jpg" /></figure><p><a href="https://blog.siray.ai/content/media/2026/11/seedance-2-5-perfume-push-in.mp4">Watch the clip</a></p><p>The camera moved straight in from the full box to a close-up of the bottle, with no drift or shake. About halfway through, a small star-shaped glint flashed on the cap, and the glass grew brighter as the camera got closer. The bottle kept its shape and stayed free of text. This is the shot to put first in a Black Friday ad.</p><h3>Shot 2: camera orbit (show the product from more than one side)</h3><pre>The camera orbits slowly around the open black gift box from left to right, about 60 degrees, staying at the height of the perfume bottle and always keeping the bottle centered. As the angle changes, the rim light moves across the amber glass and the reflection on the brushed gold cap shifts. The marble counter and the out-of-focus holiday lights slide past in parallax. The bottle, the box, the tissue paper and the burgundy ribbon do not move and keep their exact shape. Smooth gimbal move, premium product commercial. No people, no text appears.</pre><ul><li>Endpoint: seedance-2.5-i2v, 5 s, 720p, aspect_ratio: adaptive</li><li>Time: 181 seconds · Cost: $1.305 · Output: 720×1280, 24 fps, no audio track</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/0*NbxT6P1DzRtgUd9i.jpg" /></figure><p><a href="https://blog.siray.ai/content/media/2026/11/seedance-2-5-perfume-orbit.mp4">Watch the clip</a></p><p>The move was smaller than we asked for. Instead of 60 degrees, the camera made a short arc to the right, closer to 15–20 degrees. The parallax still sells it: the tree, ornaments and lanterns behind the box come into view while the bottle stays centered. To get a wider turn, ask for a full side-on view as the end point, or use the first-and-last-frame option with a second photo (see the tips below).</p><h3>Shot 3: hands-on unboxing</h3><pre>Unboxing shot. A pair of adult woman&#39;s hands, about 30 years old, with a neat nude manicure, enter from the bottom of the frame. They gently lift the amber glass perfume bottle out of the ivory tissue paper, raise it slightly toward the camera, and slowly turn it a quarter turn so the gold cap catches the light. The tissue paper rustles softly as the bottle leaves the box. The camera stays locked off at the same framing. Only the hands are visible, no face. Realistic hand anatomy with five fingers on each hand. The bottle keeps its exact shape and has no label. Warm holiday bokeh in the background. Premium commercial look. No children, teenagers, babies or minors. No text appears.</pre><ul><li>Endpoint: seedance-2.5-i2v, 5 s, 720p, aspect_ratio: adaptive</li><li>Time: 177 seconds · Cost: $1.305 · Output: 720×1280, 24 fps, no audio track</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/0*7zzvprgxAMstfLFs.jpg" /></figure><p><a href="https://blog.siray.ai/content/media/2026/11/seedance-2-5-perfume-unboxing.mp4">Watch the clip</a></p><p>This was the shot we expected to break, and it held up. Both hands came in from the bottom, with five fingers each in the frames we checked. They lifted the bottle out of the tissue and tilted it toward the camera. The bottle kept its square shape and gold cap. The “quarter turn” became more of a tilt. Saying “only the hands are visible, no face” kept the frame on the product.</p><h3>What it costs to make Black Friday video ads this way</h3><p>Seedance 2.5 i2v is billed by output size and length. Prices from the Siray AI pricing API on November 19, 2026, with a 10% discount at time of writing (list price crossed out):</p><ul><li><strong>5 seconds</strong> — 480p: $0.603 $0.67; 720p: $1.305 $1.45; 1080p: $3.199 $3.554</li><li><strong>10 seconds</strong> — 480p: $1.206 $1.34; 720p: $2.61 $2.90; 1080p: $6.397 $7.108</li><li><strong>15 seconds</strong> — 480p: $1.809 $2.01; 720p: $3.915 $4.35; 1080p: $9.596 $10.662</li><li><strong>Per second</strong> — 480p: ≈ $0.12; 720p: ≈ $0.261; 1080p: ≈ $0.64</li></ul><p>Siray AI’s pricing note gives the formula: tokens = height × width × (input duration + output duration) × 24 / 1024, at $0.012038 per 1K tokens. For our 720×1280, 5-second clip that is 108,000 tokens, or about $1.30.</p><p>Worked examples:</p><ul><li><strong>Our test set</strong> (one photo and three 5-second 720p ads): $0.035 + 3 × $1.305 = <strong>$3.95</strong>.</li><li><strong>Drafting at 480p first:</strong> three 5-second drafts cost $1.81. Re-run only the ones you like at 720p or 1080p.</li><li><strong>Ten products, three shots each, at 720p:</strong> 30 × $1.305 = $39.15, before any retakes.</li></ul><p>Check the <a href="https://console.siray.ai/model-detail/bytedance/seedance-2.5-i2v?ref=blog.siray.ai">Seedance 2.5 i2v model page</a> for live prices before you plan a batch.</p><h3>Quickstart: submit and poll</h3><p>Generation is asynchronous. Submit a task, then poll every 3–5 seconds until the status is SUCCESS.</p><p><strong>1. Submit</strong></p><pre>curl -X POST https://api.siray.ai/v1/video/generations \<br>  -H &quot;Authorization: Bearer $SIRAY_API_KEY&quot; \<br>  -H &quot;Content-Type: application/json&quot; \<br>  -d &#39;{<br>    &quot;model&quot;: &quot;bytedance/seedance-2.5-i2v&quot;,<br>    &quot;image&quot;: &quot;https://your-cdn.example.com/perfume-gift-box.jpg&quot;,<br>    &quot;prompt&quot;: &quot;Slow, smooth dolly push-in toward the amber glass perfume bottle...&quot;,<br>    &quot;duration&quot;: 5,<br>    &quot;size&quot;: &quot;720p&quot;,<br>    &quot;aspect_ratio&quot;: &quot;adaptive&quot;,<br>    &quot;audio_enable&quot;: false<br>  }&#39;</pre><p>image accepts a public URL or a data URL. We sent our 240 KB JPEG as a base64 data URL.</p><p><strong>2. Poll</strong></p><pre>curl https://api.siray.ai/v1/video/generations/&lt;task_id&gt; \<br>  -H &quot;Authorization: Bearer $SIRAY_API_KEY&quot;</pre><p>When data.status is SUCCESS, data.outputs holds the MP4 URL. Read the task_id from both the top level and data of the submit response; our earlier Siray AI tests found it at the top level. To make three shots, send three requests with the same image and different prompts. They run in parallel.</p><h3>Prompting tips for product video ads</h3><ul><li><strong>Use </strong><strong>adaptive and shape the photo, not the request.</strong> The API reference lists 9:16 as a valid aspect_ratio, but all three of our first requests with 9:16 failed with TaskTypeConstraint: &quot;For first-frame or first-last-frame generation, the output ratio follows the first-frame image.&quot; Those tasks ended as FAILURE without producing a video, and resubmitting with adaptive worked on the first try.</li><li><strong>One camera move per clip.</strong> “Push-in”, “orbit” and “locked-off” each gave a clean, readable motion. Leave multi-shot editing to your editor.</li><li><strong>List what must not move.</strong> “The bottle, the box, the tissue paper and the burgundy ribbon stay still and keep their exact shape” kept the product geometry stable in all three clips.</li><li><strong>Ask for a bigger move than you need.</strong> Our “60 degrees” orbit came out as roughly 15–20 degrees. For a wide turn, add an end_image with the product from the side.</li><li><strong>Keep offer text out of the prompt.</strong> Put “Black Friday −30%” or a price in your editor, so you can change it without paying for a new clip. Model-written text often comes out garbled.</li><li><strong>For hands, describe them precisely.</strong> Adult hands, a manicure, “five fingers on each hand”, and “only the hands are visible” gave natural results.</li></ul><h3>When to use this workflow (and when not to)</h3><p><strong>Good fit:</strong></p><ul><li>Turning an existing packshot catalog into short social ads before a sale.</li><li>Testing several ad openings (hero, orbit, unboxing) on the same product before paying for a shoot.</li><li>Gift products with clear shapes: bottles, boxes, gadgets, jewelry.</li></ul><p><strong>Not the best fit:</strong></p><ul><li><strong>Ads with on-screen prices or legal copy.</strong> Add those in post.</li><li><strong>Products with fine printed labels.</strong> i2v can blur or rewrite small text as it moves. Shoot the label as a still frame instead.</li><li><strong>Cheaper drafts:</strong> <a href="https://blog.siray.ai/kling-4-0-api-pricing-prompts/">Kling 4.0</a> image-to-video costs $0.785 for 5 seconds at 720p, and Kling 4.0 Flash costs $0.47, on Siray AI. Seedance 2.5 at 480p is $0.603.</li></ul><p>For more product ad ideas without the holiday angle, see <a href="https://blog.siray.ai/ai-product-video/">AI Product Video Ads: Turn Product Photos into Scroll-Stopping Video</a>. For how Seedance 2.5 handles film-style scenes, see <a href="https://blog.siray.ai/seedance-2-5/">Seedance 2.5 for Film &amp; TV Production</a>. For the faster Seedance tier, see the <a href="https://blog.siray.ai/seedance-2-5-flash-spicy-api-pricing-prompts/">Seedance 2.5 Flash Spicy API guide</a>.</p><h3>FAQ</h3><h3>How much does a 5-second AI product video ad cost with Seedance 2.5?</h3><p>$1.305 at 720p on Siray AI, or $0.603 at 480p and $3.199 at 1080p (prices checked November 19, 2026, with a 10% discount at time of writing). A 9:16 720p clip uses about 108,000 billable tokens.</p><h3>Can I make vertical 9:16 video ads from a product photo?</h3><p>Yes. Use a 9:16 photo and set aspect_ratio to adaptive. Seedance 2.5 image-to-video keeps the shape of the input photo. In our test, sending 9:16 explicitly was rejected.</p><h3>How long does Seedance 2.5 take to turn a photo into a video?</h3><p>Our three 5-second 720p clips took 177 to 181 seconds each, from submit to finish, while running at the same time.</p><h3>Does Seedance 2.5 add music or sound to product videos?</h3><p>It can generate audio when audio_enable is true. We set it to false. The files then had no audio track, so we could add licensed music in the editor.</p><h3>Can I put my Black Friday discount text in the prompt?</h3><p>You can, but we don’t recommend it. Generated text is often misspelled. Add prices and offer text in your editor, so one clip works for several promotions.</p><h3>Make your Black Friday video ads on Siray AI</h3><p>Upload one product photo, run the three prompts above, and you have a hero shot, an orbit and an unboxing for under $4. Start with the <a href="https://console.siray.ai/model-detail/bytedance/seedance-2.5-i2v?ref=blog.siray.ai">Seedance 2.5 i2v model page</a>, or open the <a href="https://console.siray.ai/?ref=blog.siray.ai">Siray AI console</a> to get an API key.</p><h3>Sources</h3><ul><li>ByteDance Seed, <a href="https://seed.bytedance.com/en/seedance2_5">Seedance 2.5</a>, accessed November 19, 2026.</li><li>Siray AI API reference, <a href="https://docs.siray.ai/api-reference/model-api/seedance-2.5-i2v">Seedance 2.5 i2v</a>, accessed November 19, 2026.</li><li>Siray AI API reference, <a href="https://docs.siray.ai/api-reference/model-api/seedream-5.5-t2i-spicy">Seedream 5.5 t2i Spicy</a>, accessed November 19, 2026.</li><li>Siray AI pricing API, GET /v1/models/pricing/{model}, November 19, 2026.</li></ul><p><em>Originally published on the </em><a href="https://blog.siray.ai/black-friday-product-video-ads-seedance-2-5/"><em>Siray AI blog</em></a><em>. Prices were checked on 2026–11–19 and can change; see the live model page for current pricing.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bc98d9e50e32" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Kling 4.0 API on Siray AI: Pricing, Endpoints and Tested Prompts]]></title>
            <link>https://medium.com/@siary_ai/kling-4-0-api-on-siray-ai-pricing-endpoints-and-tested-prompts-1faf1bfd1951?source=rss-81e778058314------2</link>
            <guid isPermaLink="false">https://medium.com/p/1faf1bfd1951</guid>
            <dc:creator><![CDATA[Siray AI]]></dc:creator>
            <pubDate>Wed, 18 Nov 2026 08:32:45 GMT</pubDate>
            <atom:updated>2026-11-18T08:32:45.286Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*CDIimDjUVtIDRWfC.png" /></figure><p>The Kling 4.0 API gives you Kuaishou’s Kling 4.0 video model through eight Siray AI endpoints for text-to-video, image-to-video, reference-to-video and video editing. It costs <strong>about $0.157 per second at 720p and $0.235 per second at 1080p</strong>, and the faster Kling 4.0 Flash costs $0.094 per second at 720p. It suits clips of 3 to 30 seconds with sound, from product shots to multi-shot scenes.</p><h3>Key takeaways</h3><ul><li><strong>Eight endpoints:</strong> Kling 4.0 and Kling 4.0 Flash each have t2v, i2v, ref2v and video-edit endpoints on Siray AI.</li><li><strong>Length and resolution:</strong> Kling 4.0 makes 3–30 second clips at 720p or 1080p. Flash makes 3–20 second clips at 720p only.</li><li><strong>Pricing:</strong> a 5-second clip costs $0.785 at 720p and $1.175 at 1080p with Kling 4.0, or $0.47 with Flash.</li><li><strong>Speed in our tests:</strong> Kling 4.0 jobs took 106–284 seconds. The same 5-second prompt on Flash took 86 seconds instead of 213.</li><li><strong>Audio is always on:</strong> all six of our test clips came back with a two-channel audio track, and none of the endpoints has an audio switch.</li></ul><h3>What is Kling 4.0?</h3><p>Kling 4.0 is the fourth-generation video model from Kling AI, the video product of Kuaishou Technology. Kling AI <a href="https://kling.ai/blog/kling40-kling-visual-realism-creative-control-introducing-storytelling">announced it on September 30, 2026</a>, and Kuaishou’s <a href="https://finance.yahoo.com/technology/ai/articles/kling-ai-officially-launches-next-185000472.html">press release</a> says it officially launched on October 15, 2026. Kling 4.0 is Kuaishou’s model; Siray AI offers it through its unified API.</p><p>According to the official announcement, Kling 4.0:</p><ul><li>“generates videos up to 30 seconds long in a single generation”;</li><li>accepts “up to 15 reference assets in a single generation”, split into up to 10 images, 5 videos and 7 subjects;</li><li>uses “up to 10 keyframe images to define key visual moments”;</li><li>supports “high-quality audio and two-channel stereo” with more accurate lip sync;</li><li>can “make targeted edits to character expressions, body movements, camera positions, camera movements, overall style … while keeping the rest of the scene stable”;</li><li>takes prompts of up to 8,000 tokens.</li></ul><p>Kling 4.0 Flash is the lighter tier. Kling AI describes it as “built for frequent creation” and says it “combines fast generation with excellent value”.</p><p><strong>Known limits.</strong> Kling AI’s announcement labels 10-bit HDR at 4K and video extension up to 2 minutes as “coming soon”. Kuaishou’s October 16 press release still describes 4K as upcoming. On Siray AI, the top resolution is 1080p, and the i2v endpoint takes a first frame and an optional last frame, not 10 keyframes.</p><h3>Kling 4.0 endpoints on Siray AI</h3><ul><li><strong>kuaishou/kling-4.0-t2v</strong> — Task: Text-to-video; Required inputs: prompt, duration, size, aspect_ratio</li><li><strong>kuaishou/kling-4.0-i2v</strong> — Task: Image-to-video; Required inputs: prompt, image, duration, size, aspect_ratio; Optional inputs: end_image</li><li><strong>kuaishou/kling-4.0-ref2v</strong> — Task: Reference-to-video; Required inputs: prompt, duration, size, aspect_ratio; Optional inputs: images (≤10), videos (≤5)</li><li><strong>kuaishou/kling-4.0-video-edit</strong> — Task: Edit a video; Required inputs: prompt, video, duration, size, aspect_ratio; Optional inputs: elements, keep_original_sound</li><li><strong>kuaishou/kling-4.0-flash-t2v</strong> — Task: Text-to-video; Required inputs: prompt, duration, size, aspect_ratio</li><li><strong>kuaishou/kling-4.0-flash-i2v</strong> — Task: Image-to-video; Required inputs: prompt, image, duration, size, aspect_ratio</li><li><strong>kuaishou/kling-4.0-flash-ref2v</strong> — Task: Reference-to-video; Required inputs: prompt, duration, size, aspect_ratio; Optional inputs: images (≤10), videos (≤5)</li><li><strong>kuaishou/kling-4.0-flash-video-edit</strong> — Task: Edit a video; Required inputs: prompt, video, duration, size, aspect_ratio; Optional inputs: elements, keep_original_sound</li></ul><p>Values from the Siray AI API reference:</p><ul><li>duration: any whole number from 3 to 30 on Kling 4.0, and 3 to 20 on Flash.</li><li>size: 720p or 1080p on Kling 4.0; 720p only on Flash.</li><li>aspect_ratio: auto, 16:9, 9:16, 1:1 or 21:9.</li><li>image and images accept a URL or a data URL. end_image, video and videos are documented as URLs.</li><li>elements is a list of characters or objects, each with a frontal_image, reference_images and an optional reference video.</li><li>Flash i2v has no end_image, so last-frame control needs the full model.</li></ul><h3>Kling 4.0 pricing on Siray AI</h3><p>Text-to-video, image-to-video and reference-to-video without a video input are billed per second of output:</p><ul><li><strong>5 seconds</strong> — Kling 4.0, 720p: $0.785; Kling 4.0, 1080p: $1.175; Flash, 720p: $0.47</li><li><strong>10 seconds</strong> — Kling 4.0, 720p: $1.57; Kling 4.0, 1080p: $2.35; Flash, 720p: $0.94</li><li><strong>20 seconds</strong> — Kling 4.0, 720p: $3.14; Kling 4.0, 1080p: $4.70; Flash, 720p: $1.88</li><li><strong>30 seconds</strong> — Kling 4.0, 720p: $4.71; Kling 4.0, 1080p: $7.05</li><li><strong>Per second</strong> — Kling 4.0, 720p: ≈ $0.157; Kling 4.0, 1080p: ≈ $0.235; Flash, 720p: ≈ $0.094</li></ul><p>Jobs that take a video input (video-edit, and ref2v with videos) bill both the output and the input. Siray AI&#39;s pricing note gives the formula: billable duration = output duration + max(input duration, output duration). The per-second rates for these jobs are:</p><ul><li><strong>Kling 4.0, 720p</strong> — Output: $0.094/s; Input video: $0.094/s</li><li><strong>Kling 4.0, 1080p</strong> — Output: $0.141/s; Input video: $0.141/s</li><li><strong>Flash, 720p</strong> — Output: $0.063/s; Input video: $0.063/s</li></ul><p>Worked examples:</p><ul><li><strong>A 5-second 1080p product shot:</strong> $1.175. Ten takes cost $11.75.</li><li><strong>The same 5 seconds as a Flash draft:</strong> $0.47, which is 40% less than Kling 4.0 at 720p.</li><li><strong>Editing a 5-second clip with Flash:</strong> 5 seconds out plus 5 seconds in at $0.063 each, $0.63.</li><li><strong>A full 30-second 1080p scene:</strong> $7.05.</li></ul><p>Prices were read from the Siray AI pricing API on November 17, 2026, with no discount shown. Check the <a href="https://console.siray.ai/model-detail/kuaishou/kling-4.0-t2v?ref=blog.siray.ai">Kling 4.0 t2v model page</a> for live prices. Failed requests are not billed, according to the same pricing note.</p><p><a href="https://blog.siray.ai/content/media/2026/11/kling-4-0-concept-duration.mp4">Watch the clip: a short concept video about how cost grows with clip length, made with Kling 4.0 Flash</a></p><p><em>Concept video generated with Kling 4.0 Flash on Siray AI.</em></p><h3>Quickstart: call the Kling 4.0 API</h3><p>Generation is asynchronous. Submit a task to get a task_id, then check it every 3–5 seconds. When the status is SUCCESS, download the video from data.outputs.</p><p><strong>1. Submit a task</strong></p><p>This is the request we used for our first test:</p><pre>curl -X POST https://api.siray.ai/v1/video/generations \<br>  -H &quot;Authorization: Bearer $SIRAY_API_KEY&quot; \<br>  -H &quot;Content-Type: application/json&quot; \<br>  -d &#39;{<br>    &quot;model&quot;: &quot;kuaishou/kling-4.0-t2v&quot;,<br>    &quot;prompt&quot;: &quot;A slow, low-angle dolly shot moving forward along a rain-soaked cobblestone alley...&quot;,<br>    &quot;duration&quot;: 5,<br>    &quot;size&quot;: &quot;1080p&quot;,<br>    &quot;aspect_ratio&quot;: &quot;16:9&quot;<br>  }&#39;</pre><p>In our tests the task ID came back at the top level of the response (task_id), not under data as the API reference shows, so read both places.</p><p><strong>2. Poll until the task finishes</strong></p><pre>curl https://api.siray.ai/v1/video/generations/&lt;task_id&gt; \<br>  -H &quot;Authorization: Bearer $SIRAY_API_KEY&quot;</pre><p>data.status moves through NOT_START, SUBMITTED, QUEUED and IN_PROGRESS to SUCCESS or FAILURE. On success, data.outputs holds the MP4 URL. Two details we noticed: fail_reason also contains the output URL on successful tasks, so judge success by status only, and the task record has no usage field for these endpoints.</p><p>To switch tasks, change model and add the inputs: &quot;image&quot; (and optionally &quot;end_image&quot;) for i2v, &quot;images&quot;: [...] for ref2v, or &quot;video&quot;: &quot;&lt;URL&gt;&quot; for video-edit.</p><p><a href="https://blog.siray.ai/content/media/2026/11/kling-4-0-concept-async-flow.mp4">Watch the clip: a short concept video about the submit-and-poll flow, made with Kling 4.0 Flash</a></p><p><em>Concept video generated with Kling 4.0 Flash on Siray AI.</em></p><h3>Kling 4.0 prompt examples (tested)</h3><p>We ran every example below on November 17, 2026 through the Siray AI API. Times are measured from submit to finish on the task record. Costs are calculated from Siray AI’s list prices, because the task record does not report usage for these endpoints. Every clip came back at 24 fps with a two-channel audio track.</p><h3>1. Camera move at 1080p</h3><pre>A slow, low-angle dolly shot moving forward along a rain-soaked cobblestone alley in a Japanese mountain town at dusk. Paper lanterns glow above wooden shopfronts, steam drifts from a small noodle stall, and puddles reflect the warm light. In the final second the camera rises smoothly above the rooftops to reveal forested mountains under a deep blue sky. Cinematic 35mm film look, soft grain, natural motion. No people. No text, no signs with letters, no logos.</pre><ul><li>Endpoint: kling-4.0-t2v, 5 s, 1080p, 16:9</li><li>Time: 213 seconds · Cost: $1.175 · Output: 1920×1080</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mPYeaKywOApFaqAC.jpg" /></figure><p><a href="https://blog.siray.ai/content/media/2026/11/kling-4-0-alley-dolly-1080p.mp4">Watch the clip</a></p><p>The low dolly, the steam, the puddle reflections and the final rise over the rooftops all appear, in that order. Putting the camera move in a sentence of its own (“In the final second the camera rises…”) gave the model a clear point to change the motion.</p><h3>2. The same prompt on Kling 4.0 Flash</h3><ul><li>Endpoint: kling-4.0-flash-t2v, 5 s, 720p, 16:9, same prompt as example 1</li><li>Time: 86 seconds · Cost: $0.47 · Output: 1280×720</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*HmK_NtszVBhTNYVg.jpg" /></figure><p><a href="https://blog.siray.ai/content/media/2026/11/kling-4-0-flash-alley-dolly-720p.mp4">Watch the clip</a></p><p>Flash finished in less than half the time and kept the dolly and the rise. The look changed more than the motion did. Flash replaced the scattered lanterns with long, even rows, smoothed the cobblestones and made the scene cleaner and more stylized. It works well for checking timing and camera moves before you pay for a 1080p take.</p><h3>3. Three shots in one 10-second clip</h3><pre>A three-shot sequence in a small bicycle repair shop on a sunny morning, natural documentary look.<br>Shot 1 (wide): a mechanic in his 40s with a short grey beard and a green canvas apron rolls up the metal shutter of the shop and wheels a blue bicycle inside.<br>Shot 2 (close-up): his hands fit a new chain onto the rear sprocket and spin the pedal; the wheel turns smoothly.<br>Shot 3 (medium shot): he hands the bicycle to a woman in her 30s at the door, she nods, and he wipes his hands on a rag and smiles.<br>Keep the same man, apron, bicycle and shop in every shot. Adults only, no children, teenagers, babies or minors. No text, no logos.</pre><ul><li>Endpoint: kling-4.0-t2v, 10 s, 720p, 16:9</li><li>Time: 236 seconds · Cost: $1.57 · Output: 1280×720</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*B4-rvhfPFJ_gBqtF.jpg" /></figure><p><a href="https://blog.siray.ai/content/media/2026/11/kling-4-0-bike-shop-multi-shot.mp4">Watch the clip</a></p><p>Kling 4.0 cut between three distinct shots and kept the same man, apron, blue bicycle and shop throughout. Labelling each shot with its framing (wide, close-up, medium) gave clean cuts. The close-up showed the rear sprocket and chain, though the hands turned the sprocket rather than spinning the pedal.</p><h3>4. Image-to-video with a first and last frame</h3><p>We gave the i2v endpoint two photos of the same woman on a rooftop. The image shows her in a moss-green wool coat. The end_image shows her in just the cream turtleneck and charcoal trousers, with the coat folded on the wall behind her. We made the end frame by editing the first photo with Seedream 5.5, so the face, light and background match exactly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*EylxYK4dQmpha7xY.jpg" /></figure><pre>The woman on the rooftop slowly unbuttons her moss-green wool coat, slips it off her shoulders, folds it once and lays it on the white parapet wall to her right, leaving her in the cream turtleneck sweater and charcoal trousers. She then turns back toward the camera and settles into the same pose as the end frame. Warm golden-hour light, gentle breeze in her hair, steady handheld camera. Adult woman only, fully clothed. No text, no logos.</pre><ul><li>Endpoint: kling-4.0-i2v with end_image, 5 s, 720p, aspect_ratio: auto</li><li>Time: 284 seconds · Cost: $0.785 · Output: 836×1116, matching the shape of the input photos</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/836/0*mm4jOBTLwCyPJWfd.jpg" /></figure><p><a href="https://blog.siray.ai/content/media/2026/11/kling-4-0-i2v-coat-off.mp4">Watch the clip</a></p><p>The clip starts on the first photo and lands on the end frame, with the same face, sweater, rooftop and light. In between, Kling 4.0 played the whole action in order: she unbuttons the coat, slides it off both shoulders, turns away to lay it on the wall and turns back to the camera. The path is approximate, not exact. For about a second the coat stands up stiffly on the wall before it settles into the folded shape from the end frame.</p><p>An earlier version of this test used a navy blazer of the same cut as the end frame. That made the result read as a color swap rather than a coat coming off. <strong>An end frame that differs clearly in shape, not only in color, gives the model a transition it can show.</strong></p><h3>5. Reference-to-video: a person and a product</h3><p>We passed two images to ref2v: a potter in her studio and a stoneware pour-over coffee set on a plain background.</p><pre>Use image 1 as the location and the person, and image 2 as the product. The grey-haired potter from image 1 stops her wheel, wipes her hands on her apron and walks to a wooden table by the rainy window, where the stoneware pour-over coffee set from image 2 is waiting. She pours hot water from the kettle into the dripper in a slow spiral while steam rises in the soft window light. The camera pushes in slowly from a medium shot to a close-up of the coffee dripping into the cup. Keep the shape, glaze and colour of the coffee set exactly as in image 2, and keep the potter&#39;s face, hair and apron as in image 1. Adults only. No text, no logos.</pre><ul><li>Endpoint: kling-4.0-ref2v, images with 2 items, 5 s, 720p, 16:9</li><li>Time: 106 seconds · Cost: $0.785 · Output: 1280×720</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YdBotCzz0S--Lv3H.jpg" /></figure><p><a href="https://blog.siray.ai/content/media/2026/11/kling-4-0-ref2v-potter-coffee.mp4">Watch the clip</a></p><p>The product held up best: the fluted dripper, the brass ring and the spouted pitcher match the reference photo. The studio, the rainy window and the potter’s hair and apron also carried over. The opening was compressed: she is already walking with the kettle when the clip begins, so the wheel and the hand wipe are skipped. Numbering the images (“image 1”, “image 2”) and saying what each is for kept the two references apart.</p><h3>6. Video edit with Flash: rain to snow</h3><p>We sent the Flash clip from example 2 to the Flash video-edit endpoint.</p><pre>Change the weather from rain to a quiet snowfall. Cover the cobblestones, rooftops and the noodle stall roof with a fresh layer of white snow, replace the puddles with soft snow, and add slow, large snowflakes falling through the lantern light. Keep the same street, buildings, lanterns, steam and camera movement as the original video. No people. No text, no logos.</pre><ul><li>Endpoint: kling-4.0-flash-video-edit, 5 s input, 5 s, 720p, 16:9</li><li>Time: 74 seconds · Cost: $0.63 (5 s output + 5 s input) · Output: 1280×720</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*pwzw0e7k4wuoFgTJ.jpg" /></figure><p><a href="https://blog.siray.ai/content/media/2026/11/kling-4-0-flash-video-edit-snow.mp4">Watch the clip</a></p><p>The street, lanterns, steam, stall and camera move stayed the same, and snow now covers the ground and roofs. One detail went the opposite way: the spots where the puddles were became bare patches of cobblestone in the snow. We could not see falling snowflakes in the frames we checked. This was the fastest job in our tests.</p><h3>Prompting tips for Kling 4.0</h3><ul><li><strong>Give each camera move its own sentence.</strong> “In the final second the camera rises…” produced a clean change of motion at the right moment.</li><li><strong>Label shots with their framing.</strong> “Shot 1 (wide) … Shot 2 (close-up) …” gave three clear cuts in a 10-second clip. Repeat what must stay the same in every shot.</li><li><strong>Number reference images and give each a role.</strong> “Use image 1 as the location and the person, and image 2 as the product” kept the two inputs apart.</li><li><strong>Put the key action early.</strong> In both image-driven tests, the first step of a multi-step action was the one most likely to be skipped or shortened.</li><li><strong>Draft on Flash, finish on Kling 4.0.</strong> The Flash take kept the camera move at 40% of the 720p price, so you can check timing before paying for 1080p.</li><li><strong>Plan for sound.</strong> Every clip comes with audio. Describe the sound you want, or strip the track in editing if you don’t need it.</li></ul><h3>When to use Kling 4.0 (and when not to)</h3><p><strong>Use Kling 4.0 for:</strong></p><ul><li>Final 1080p clips up to 30 seconds with sound.</li><li>Multi-shot scenes where the same person and props must carry across cuts.</li><li>Product videos from reference photos, where the product’s shape and finish must hold.</li><li>First-and-last-frame transitions, such as outfit or product reveals.</li></ul><p><strong>Use Kling 4.0 Flash for:</strong></p><ul><li>Drafts, timing checks and A/B variants at 720p.</li><li>High-volume social clips up to 20 seconds.</li><li>Quick video edits, such as changing weather or time of day.</li></ul><p><strong>Look elsewhere when:</strong></p><ul><li>You need 4K or clips longer than 30 seconds: Kling AI lists 4K and 2-minute extension as coming soon, and Siray AI does not offer them yet.</li><li>You need control over every frame with up to 10 keyframes: the Siray AI i2v endpoint takes a first and a last frame.</li><li>You need a lower price per second at 480p: <a href="https://blog.siray.ai/seedance-2-5-flash-spicy-api-pricing-prompts/">Seedance 2.5 Flash</a> costs about $0.068 per second at 480p.</li></ul><p>For more Kling examples, see <a href="https://blog.siray.ai/kling-3-0-omni/">Kling 3.0 Omni</a>, <a href="https://blog.siray.ai/kling-3-motion-control/">Kling 3.0 Motion Control</a> and the comparison <a href="https://blog.siray.ai/veo-vs-kling-vs-seedance-2-0/">Veo vs Kling vs Seedance 2.0</a>.</p><h3>FAQ</h3><h3>How much does the Kling 4.0 API cost per second?</h3><p>About $0.157 per second at 720p and $0.235 per second at 1080p on Siray AI. A 5-second clip costs $0.785 at 720p and $1.175 at 1080p (prices checked November 17, 2026).</p><h3>What is the difference between Kling 4.0 and Kling 4.0 Flash?</h3><p>Flash is the faster, cheaper tier. On Siray AI it outputs 720p only, makes clips of up to 20 seconds instead of 30, and its i2v endpoint has no last frame. It costs about $0.094 per second, and in our test the same prompt finished in 86 seconds instead of 213.</p><h3>How long can a Kling 4.0 video be?</h3><p>From 3 to 30 seconds on Kling 4.0 and 3 to 20 seconds on Flash, chosen with duration.</p><h3>Does Kling 4.0 generate audio?</h3><p>Yes. Every clip in our tests came with a two-channel audio track. None of the Siray AI endpoints has a switch to turn audio off. On the video-edit endpoints, keep_original_sound keeps the input video&#39;s sound.</p><h3>Can Kling 4.0 edit an existing video?</h3><p>Yes. Send a video URL to kuaishou/kling-4.0-video-edit or kuaishou/kling-4.0-flash-video-edit with a prompt describing the change. You pay for both the output and the input duration. Our 5-second Flash edit cost $0.63.</p><h3>How long does Kling 4.0 take to generate a video?</h3><p>In our tests, Kling 4.0 took 106 to 284 seconds for 5- to 10-second clips. Flash took 74 to 86 seconds.</p><h3>Try Kling 4.0 on Siray AI</h3><p>Get an API key and run your first clip from the <a href="https://console.siray.ai/model-detail/kuaishou/kling-4.0-t2v?ref=blog.siray.ai">Kling 4.0 t2v model page</a>. The same key works for all eight Kling 4.0 endpoints and every other model on <a href="https://console.siray.ai/?ref=blog.siray.ai">Siray AI</a>.</p><h3>Sources</h3><ul><li>Kling AI, <a href="https://kling.ai/blog/kling40-kling-visual-realism-creative-control-introducing-storytelling">Kling 4.0: visual realism, creative control and storytelling</a>, September 30, 2026, accessed November 17, 2026.</li><li>Kuaishou Technology, <a href="https://finance.yahoo.com/technology/ai/articles/kling-ai-officially-launches-next-185000472.html">Kling AI officially launches next-generation model</a>, press release, accessed November 17, 2026.</li><li>Siray AI API reference: <a href="https://docs.siray.ai/api-reference/model-api/kling-4.0-t2v">t2v</a>, <a href="https://docs.siray.ai/api-reference/model-api/kling-4.0-i2v">i2v</a>, <a href="https://docs.siray.ai/api-reference/model-api/kling-4.0-ref2v">ref2v</a>, <a href="https://docs.siray.ai/api-reference/model-api/kling-4.0-video-edit">video-edit</a>, <a href="https://docs.siray.ai/api-reference/model-api/kling-4.0-flash-t2v">Flash t2v</a>, <a href="https://docs.siray.ai/api-reference/model-api/kling-4.0-flash-i2v">Flash i2v</a>, <a href="https://docs.siray.ai/api-reference/model-api/kling-4.0-flash-ref2v">Flash ref2v</a>, <a href="https://docs.siray.ai/api-reference/model-api/kling-4.0-flash-video-edit">Flash video-edit</a>, accessed November 17, 2026.</li><li>Siray AI pricing API, GET /v1/models/pricing/{model}, November 17, 2026.</li></ul><p><em>Originally published on the </em><a href="https://blog.siray.ai/kling-4-0-api-pricing-prompts/"><em>Siray AI blog</em></a><em>. Prices were checked on 2026–11–17 and can change; see the live model page for current pricing.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1faf1bfd1951" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Nano Banana Pro API on Siray AI: Pricing, 4K Output and Tested Prompts]]></title>
            <link>https://medium.com/@siary_ai/nano-banana-pro-api-on-siray-ai-pricing-4k-output-and-tested-prompts-0420ee5b19c3?source=rss-81e778058314------2</link>
            <guid isPermaLink="false">https://medium.com/p/0420ee5b19c3</guid>
            <dc:creator><![CDATA[Siray AI]]></dc:creator>
            <pubDate>Wed, 18 Nov 2026 05:45:30 GMT</pubDate>
            <atom:updated>2026-11-18T05:45:30.084Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*HwSassw2q7p0SWvp.png" /></figure><p>The Nano Banana Pro API gives you Google’s Gemini 3 Pro Image model for generation and multi-reference editing. On Siray AI it costs <strong>$0.098 per 2K image and $0.168 per 4K image</strong>, with up to 14 input images charged at no extra cost. It is the strongest model we have tested at rendering real paragraphs of text, and the weakest we have tested at getting numbers inside a chart right. This guide covers the endpoints, pricing, a quickstart and seven images we generated end to end.</p><h3>Key takeaways</h3><ul><li><strong>Text is genuinely solved, at paragraph length.</strong> A four-line body paragraph, a headline, a subhead and a footer line came back character-perfect in one 2K image. The same held at 4K, down to a line of type about 12 pixels tall.</li><li><strong>14 reference images cost the same as one.</strong> $0.098 for a three-reference composite and $0.098 for fourteen. That is the opposite of Seedream 5.5, which charges $0.008 per input image.</li><li><strong>4K is real 4K.</strong> A 16:9 render came back at 5504x3072, double the 2K frame on each axis — but 19–26 MB per file.</li><li><strong>It draws charts it cannot count.</strong> Our infographic came out beautifully typeset with three timeline bars whose lengths contradicted their own labels. Never ship a diagram from this model without checking the numbers.</li><li><strong>There is no </strong>output_format<strong> field.</strong> Output is always PNG, so plan for file sizes rather than choosing a cheaper format.</li><li><strong>Fast for what it is.</strong> Every job finished in 31–46 seconds at 2K and 4K; the 14-reference edit took 21.</li></ul><h3>What is Nano Banana Pro?</h3><p>Nano Banana Pro is Google’s product name for <strong>Gemini 3 Pro Image</strong>, announced on <a href="https://blog.google/innovation-and-ai/products/nano-banana-pro/?ref=blog.siray.ai">the Google blog on November 20, 2025</a> by Naina Raisinghani of Google DeepMind. It is built on Gemini 3 Pro and sits above the original Nano Banana, which Google identifies as the Gemini 2.5 Flash Image model.</p><p>Google’s stated strengths:</p><ul><li><strong>Text rendering.</strong> Google calls it “the best model for creating images with correctly rendered and legible text directly in the image”, covering both a short tagline and a long paragraph, with a wider range of textures, fonts and calligraphy. Multilingual reasoning lets it write or localise copy inside an image.</li><li><strong>Resolution.</strong> Output is “available 2K and 4K resolution” across a range of aspect ratios pitched at social and print work.</li><li><strong>Reference images.</strong> It can “blend more elements than ever before, using up to 14 images” while “maintaining the consistency and resemblance of up to 5 people”.</li><li><strong>Editing controls.</strong> Improved localised editing, plus camera-angle changes, focus and depth-of-field shifts, colour grading and relighting.</li><li><strong>Reasoning and world knowledge.</strong> The model “uses Gemini’s state-of-the-art reasoning and real-world knowledge to visualize information better than ever before”, and in Google’s own products can ground on Google Search.</li></ul><p><strong>What Google does not say.</strong> The announcement is promotional and publishes no error rates, latency figures or quality benchmarks. It concedes only that free-tier users fall back to the original Nano Banana after a quota, and that all output carries Google’s imperceptible SynthID watermark, with a visible Gemini sparkle on free and AI Pro tiers. It frames Pro as a trade-off rather than a pure upgrade — the original is for “fast, fun editing”, Pro for “complex compositions requiring the highest quality” — which implies cost and speed as limits. There is no failure-mode section at all.</p><h3>Nano Banana Pro endpoints on Siray AI</h3><p>Siray AI exposes two endpoints, both non-Spicy:</p><p>Model IDTaskRequired inputsNotesgoogle/nano-banana-pro-t2iText-to-imageprompt, size, aspect_ratioNo output_format field; output is PNGgoogle/nano-banana-pro-i2iEditing and multi-reference compositingprompt, size, aspect_ratio, images (1–14)Input images are URLs or data URLs, and are not billed separately</p><p>size takes only three values — 1k, 2k, 4k. The ratio comes from aspect_ratio, not from the size string, which is a different design from Seedream and FLUX.</p><p>ParameterValuessize1k, 2k, 4kaspect_ratio1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9</p><p>Both parameters are required, including aspect_ratio — there is no default. No output_format, quality, seed, negative_prompt or num_images field exists on either endpoint, so each request returns exactly one image.</p><h3>Nano Banana Pro pricing on Siray AI</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MUfKT1VTuMFm59g5.png" /></figure><p><em>Illustration generated with Nano Banana Pro on Siray AI.</em></p><p>TierPriceList price1K$0.098$0.142K$0.098$0.144K$0.168$0.24</p><p>A 30% discount was active when we checked on November 12, 2026. Both endpoints cost the same, so editing is no more expensive than generating — and <strong>the number of input images does not change the price</strong>.</p><p>That last point is worth spelling out, because it is the opposite of how most editing endpoints on Siray bill:</p><p>RequestNano Banana ProSeedream 5.5 (for comparison)Edit with 1 reference image$0.098$0.043Edit with 3 reference images$0.098$0.059Edit with 14 reference images$0.098not supported (max 10)</p><p>Worked examples:</p><ul><li><strong>1,000 social images at 2K:</strong> 1,000 × $0.098 = <strong>$98</strong>.</li><li><strong>The same 1,000 images at 4K:</strong> 1,000 × $0.168 = <strong>$168</strong>.</li><li><strong>A 14-reference brand composite:</strong> <strong>$0.098</strong>, regardless of how many references you pack in.</li></ul><p>Prices were read from the Siray AI pricing API on November 12, 2026. Check the <a href="https://console.siray.ai/model-detail/google/nano-banana-pro-t2i?ref=blog.siray.ai">Nano Banana Pro model page</a> for live prices.</p><h3>Quickstart: call the Nano Banana Pro API</h3><p>Generation is asynchronous. Submit a task, get a task_id, then poll every 3–5 seconds until the status is SUCCESS. Image URLs are in data.outputs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*IftjBi48YrmytpLx.png" /></figure><p><em>Illustration generated with Nano Banana Pro on Siray AI.</em></p><p><strong>1. Submit a task</strong></p><pre>curl -X POST https://api.siray.ai/v1/images/generations/async \<br>  -H &quot;Authorization: Bearer $SIRAY_API_KEY&quot; \<br>  -H &quot;Content-Type: application/json&quot; \<br>  -d &#39;{<br>    &quot;model&quot;: &quot;google/nano-banana-pro-t2i&quot;,<br>    &quot;prompt&quot;: &quot;A vertical concert poster for a solo cellist, charcoal background, cream condensed serif type&quot;,<br>    &quot;size&quot;: &quot;4k&quot;,<br>    &quot;aspect_ratio&quot;: &quot;2:3&quot;<br>  }&#39;</pre><p><strong>2. Poll for the result</strong></p><pre>curl https://api.siray.ai/v1/images/generations/async/$TASK_ID \<br>  -H &quot;Authorization: Bearer $SIRAY_API_KEY&quot;</pre><p>The status moves through SUBMITTED or QUEUED and IN_PROGRESS to SUCCESS or FAILURE. We saw QUEUED rather than SUBMITTED on one 4K job, so handle both. On failure, read data.fail_code and data.fail_reason. Full parameter details are in the <a href="https://docs.siray.ai/api-reference/model-api/nano-banana-pro-t2i?ref=blog.siray.ai">Nano Banana Pro t2i API reference</a>.</p><p><strong>Three things we found in testing that the reference does not spell out:</strong></p><ol><li><strong>Output is always PNG, and there is no field to change it.</strong> The schema has no output_format, and every one of our seven images came back as PNG. A 4K landscape was 26 MB. Budget storage and bandwidth accordingly; there is no JPEG escape hatch on this endpoint.</li><li>aspect_ratio<strong> is required, with no default.</strong> Omitting it is a validation error rather than a silent 1:1. If you are porting code from another Siray image model that picks a ratio from a size string like 2816x1584, you have to rewrite that part.</li><li><strong>The number of input images does not affect billing.</strong> Three references and fourteen both cost $0.098. Nothing in the response or the docs confirms this — we verified it by comparing two runs on the same day.</li></ol><h3>Nano Banana Pro prompt examples (tested)</h3><p>We ran every prompt below through the Siray AI API on November 12, 2026. Timings include queueing. Seven images cost $0.826 in total: five text-to-image and two editing runs.</p><h3>1. A poster with a real body paragraph</h3><pre>A single poster for an independent coffee roastery, portrait orientation. Background is a flat deep aubergine. Top third: a small letterpress monogram &quot;HR&quot; in a thin serif, hairline rule beneath it. Middle: the headline &quot;HARVEST LANE ROASTERS&quot; set in large condensed all-caps serif, two lines, generous letter spacing. Below it a smaller subhead in italic: &quot;Single origin, roasted in Bermondsey&quot;. Lower third: a short paragraph of small body text, four lines, reading exactly: &quot;We buy two harvests a year and roast them the week they land. Ask at the counter for the current lot and we will grind it to whatever you brew with.&quot; At the base a thin rule, then in small capitals: &quot;EST. 2019 · 14 HARVEST LANE · OPEN TUES TO SAT 7-4&quot;. Print-flat colour, no photography, no gradients, precise typography, even optical margins, all letterforms perfectly formed with no misspellings and no stray characters.</pre><p>Settings: text-to-image, size: &quot;2k&quot;, aspect_ratio: &quot;3:4&quot;, 31 seconds, $0.098. Output 1792x2400.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*bLiknrILibL_yIv2.png" /></figure><p>This is the strongest text result we have seen from any image model. The four-line paragraph is character-perfect, including the apostrophe-free “we will” and the hyphenated date range in the footer. Letterforms are consistent between the headline and the body, and the baseline spacing in the paragraph is even. Most models that handle a headline will garble a paragraph this long; this one did not.</p><h3>2. Product photography with real material texture</h3><pre>Editorial product photograph of a pair of dark olive suede desert boots standing on a pale limestone ledge, three-quarter angle, both boots fully in frame. Visible suede nap, crepe sole with a thin rim of dirt, waxed laces threaded through four brass eyelets, faint crease where the toe flexes. Soft directional daylight from the upper left with a long soft shadow falling right, neutral warm-grey background with a gentle falloff, shallow depth of field with both the near boot and the far boot sharp. Sharp focus on stitching and eyelets, no skin, no hands, no logos, no text.</pre><p>Settings: text-to-image, size: &quot;2k&quot;, aspect_ratio: &quot;3:2&quot;, 41 seconds, $0.098. Output 2528x1696.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eYtUhAhKwOOPqEzv.png" /></figure><p>Followed the brief closely: both boots in frame, the suede nap reads as real material rather than a painted texture, and the shadow falls in the direction asked for. The one thing it invented is the crepe sole texture, which is chunkier than the reference footwear it is imitating. Nothing is misspelled because the prompt asked for no text — worth noting, since models that specialise in text sometimes add stray lettering to product shots.</p><h3>3. An infographic that got the layout right and the data wrong</h3><pre>A clean editorial infographic poster titled &quot;HOW A POUR-OVER BREWS&quot;. Landscape. Left column: an exploded diagram of a paper filter cone sitting on a glass carafe, drawn as flat vector line art with thin consistent strokes, labelled with small leader lines to exactly three callouts: &quot;PAPER FILTER&quot;, &quot;16G MEDIUM GRIND&quot;, &quot;300ML WATER AT 94C&quot;. Right column: three stacked horizontal bars showing a timeline, each bar a different length, labelled in small caps &quot;BLOOM 40S&quot;, &quot;FIRST POUR 1M 10S&quot;, &quot;DRAW DOWN 3M&quot;. Along the bottom a single line of small print reading exactly: &quot;Grind coarse. Pour slow. Wait.&quot; Palette is bone white background with one accent colour, burnt orange, used only for the water and the timeline bars. Precise aligned typography, all text correctly spelled, no decorative clutter.</pre><p>Settings: text-to-image, size: &quot;2k&quot;, aspect_ratio: &quot;16:9&quot;, 31 seconds, $0.098. Output 2752x1536.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zirtNjX9nfv0jpfg.png" /></figure><p><strong>This is the example to study.</strong> The layout is genuinely good: consistent stroke weight, three callouts with leader lines that point where they should, a restrained two-colour palette, and every label spelled correctly.</p><p>The data is nonsense. All three timeline axes end at “4M 50S”, but the bar labelled “BLOOM 40S” runs about a sixth of the axis, “FIRST POUR 1M 10S” about half, and “DRAW DOWN 3M” the full width. Read literally, the poster says a three-minute step takes four minutes and fifty seconds, and a forty-second bloom takes it seven times over. The model typeset a chart without computing it.</p><p>The practical rule: <strong>use it for the layout, not the arithmetic.</strong> Generate the frame, then redraw the bars in Figma or Illustrator against your real data. For diagrams where only the labels matter, it is a large time saving. For anything a reader might measure, it is not safe to ship.</p><h3>4–5. 2K versus 4K</h3><p>We ran the same class of prompt at both tiers to see what the resolution step buys.</p><pre>Aerial landscape photograph of a tidal salt marsh at low tide, looking down at a shallow braided channel winding through olive-green samphire beds. Mud flats show fine drainage rilles and small pooled water reflecting a pale overcast sky. A single narrow wooden footbridge crosses the channel at the upper third. Soft diffuse light, no hard shadows, muted natural greens and ochres, high micro-contrast in the mud texture, extremely fine detail in the vegetation edges, no people, no boats, no text, no logos.</pre><p>Settings: text-to-image, size: &quot;4k&quot;, aspect_ratio: &quot;16:9&quot;, 46 seconds, $0.168. Output <strong>5504x3072</strong>.</p><p>The 4K tier is not an upscale of a smaller render. Drainage rilles in the mud hold their edges at 100%, the flooded pools carry distinct reflection gradients, and the samphire clumps have individual leaf texture at their borders rather than smearing into green mush. This is the kind of landscape stock work where the extra $0.07 over 2K is easy to justify.</p><pre>A vertical concert poster for a solo cellist, portrait orientation, printed on heavy uncoated stock photographed flat. Background is a dense charcoal-black with a faint paper grain. Upper half: a single high-contrast duotone photograph of a cello&#39;s f-hole and strings, cropped tight, rendered only in cream and black. Lower half: the artist name &quot;MIRA OKONKWO&quot; in very large condensed serif capitals on two lines, then beneath it in small letter-spaced capitals &quot;SOLO CELLO · THREE NIGHTS&quot;, then a hairline rule, then four lines of small centred text reading exactly: &quot;ST CUTHBERT&#39;S HALL&quot;, &quot;14-16 NOVEMBER 2026&quot;, &quot;DOORS 19:30 · NO SUPPORT&quot;, &quot;TICKETS FROM 18 GBP&quot;. At the very base, tiny centred text: &quot;PRINTED IN AN EDITION OF 300&quot;. Strict grid, generous margins, all type correctly spelled and optically aligned, no decorative flourish.</pre><p>Settings: text-to-image, size: &quot;4k&quot;, aspect_ratio: &quot;2:3&quot;, 41 seconds, $0.168. Output <strong>3392x5056</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fk8DYl6ql7isjZTd.png" /></figure><p>Every character is correct, including the hyphens in “14–16”, the middle dots, the “19:30” colon and the 12-pixel “PRINTED IN AN EDITION OF 300” at the base. The duotone treatment is applied to the instrument only and does not bleed into the type. For print work this is close to shippable as-is — the only correction most designers would make is optical kerning on the artist name.</p><h3>6. Three reference images, one brand system</h3><pre>Use the three reference images as one brand system. Build a single new image: a landscape banner for the same coffee roastery, using the aubergine background and serif typography from the poster, the same warm neutral light and limestone surface from the product photograph, and the same burnt-orange accent and thin-line diagram language from the infographic. Centre the composition on a small pour-over dripper on the limestone ledge at left, with the headline &quot;HARVEST LANE ROASTERS&quot; in the poster&#39;s condensed serif capitals at right, and beneath it a single small line in the poster&#39;s italic: &quot;Single origin, roasted in Bermondsey.&quot; Add one thin burnt-orange rule under the headline, the only other colour in the frame. Keep vast empty aubergine space above the headline. All text correctly spelled, one single flat print finish, no collage edges.</pre><p>Settings: image-to-image, 3 input images, size: &quot;2k&quot;, aspect_ratio: &quot;16:9&quot;, 10 seconds, $0.098.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*srx3agciBWUQiWT7.png" /></figure><p>This is what the model is actually for. It took the aubergine from one reference, the limestone and light from another, the orange accent from the third, and produced one flat poster with no visible collage seam and a perfectly spelled headline and subhead. Ten seconds for a three-way brand composite at $0.098 is the cheapest competent version of this job we have tested.</p><h3>7. All fourteen references at once</h3><pre>Build one contact sheet from all fourteen reference crops. The sheet is a landscape grid, seven columns by two rows, on a bone-white paper background with a thin hairline rule between every cell. Each cell holds one crop, uncropped further, fitted inside its cell with consistent gutters. Above the grid, a single line of small letter-spaced capitals reading exactly &quot;HARVEST LANE ROASTERS · VISUAL ARCHIVE&quot;. Below the grid, a single line of small print reading exactly &quot;Fourteen frames, one palette.&quot; Do not add captions inside the cells. Unify every crop under one consistent warm neutral grade so the sheet reads as a single printed page, no collage edges, no drop shadows and no duplicated frames.</pre><p>Settings: image-to-image, 14 input images, size: &quot;2k&quot;, aspect_ratio: &quot;16:9&quot;, 21 seconds, $0.098.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*OmiNeR8jqAMpxaWu.png" /></figure><p>Fourteen is the documented maximum and it did not drop a frame: all fourteen crops appear, once each, in the right order, under the requested heading, with the requested footer line spelled correctly. The header and footer are the cleanest text in the image.</p><p>The problem is density. Squeezing fourteen sources into a 7×2 grid on a 16:9 canvas forces each cell to crop its source further, so text inside the references gets cut — “HARVEST LANE” reads as “ARVEST LAN”, “MIRA OKONKWO” as “KONKWO”. On a 2K canvas each cell is roughly 390 px wide. If your references contain legible type, go to 4K or give the grid fewer cells.</p><h3>Prompting tips for Nano Banana Pro</h3><ol><li><strong>Quote every string exactly, in the order you want it.</strong> The model reproduces quoted text reliably; the ones that came out wrong were the ones we did not quote.</li><li><strong>Say “reading exactly” before long copy.</strong> Our four-line paragraph and our four-line concert details were both introduced that way, and both came back character-perfect.</li><li><strong>Never trust numbers it has to compute.</strong> Bar lengths, axis scales, percentages and totals are typeset, not calculated. Redraw any chart you intend to publish.</li><li><strong>Use references as a mood board, not a paste-up.</strong> Describing what each reference contributes — “the aubergine from the poster, the limestone from the product shot” — is what produced the seamless three-way composite.</li><li><strong>Budget cells, not references, at high counts.</strong> Fourteen inputs cost no more than one, but each cell shrinks, and small type inside the references stops being legible. Go 4K above about six references.</li><li><strong>Ask for a flat print finish.</strong> “One single flat print finish, no collage edges” stopped the model from compositing visible seams.</li><li><strong>Remember there is no JPEG option.</strong> Every output is PNG. A 4K render can be 19–26 MB, which matters if you are serving images directly.</li></ol><h3>When to use Nano Banana Pro (and when not to)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*HjxIIhPJOiLPh66t.png" /></figure><p><em>Illustration generated with Nano Banana Pro on Siray AI.</em></p><p><strong>Good fit:</strong></p><ul><li><strong>Print and poster work with real copy.</strong> Long paragraphs, small legal lines, multi-language localisation of existing artwork.</li><li><strong>Multi-reference brand composites.</strong> Three to six references at $0.098 flat, with typography that survives.</li><li><strong>4K hero images.</strong> Genuine 5504x3072 output with texture that holds up in print.</li><li><strong>Product and landscape photography</strong> where material realism matters more than style.</li></ul><p><strong>Consider another model when:</strong></p><ul><li><strong>You need numbers that add up.</strong> Use it for the frame, then rebuild the chart. There is no prompt that fixes this.</li><li><strong>You are generating at volume.</strong> At $0.098 per 2K image it is roughly five times the cost of <a href="https://blog.siray.ai/seedream-5-5-api-pricing-prompts/">Seedream 5.5</a> at $0.02. For thumbnails and iteration, 5.5 is the better buy.</li><li><strong>You need many editing rounds on one image.</strong> <a href="https://blog.siray.ai/consistent-outfit-changes-with-seedream-5-0-pro-spicy-i2i/">Seedream 5.5’s editing chain</a> is cheaper per round and we have tested it to five rounds without drift.</li><li><strong>You want fast, cheap variations rather than one considered image.</strong> The original Nano Banana, or <a href="https://blog.siray.ai/z-image-turbo/">Z-Image</a>, fits that job better.</li><li><strong>You need a JPEG output</strong> to keep file sizes down. This endpoint does not offer one.</li></ul><h3>FAQ</h3><h3>How much does the Nano Banana Pro API cost?</h3><p>On Siray AI, $0.098 per 1K or 2K image and $0.168 per 4K image, on both the text-to-image and editing endpoints. A 30% discount was active on November 12, 2026, against a list price of $0.14 and $0.24. Input images are not charged separately, so a 14-reference composite costs the same as a single-image edit.</p><h3>Does Nano Banana Pro charge per reference image?</h3><p>No. On Siray AI the editing endpoint is a flat $0.098 at 2K or $0.168 at 4K regardless of whether you pass 1 or 14 images. That is different from Seedream 5.5, which adds $0.008 for every input image.</p><h3>What resolution does Nano Banana Pro output?</h3><p>Three tiers: 1k, 2k and 4k, chosen with the size parameter, with the shape set separately by aspect_ratio. In our tests a 16:9 4K render came back at 5504x3072 and a 2:3 4K render at 3392x5056, so 4K is roughly double the 2K frame on each axis rather than an upscale.</p><h3>Can Nano Banana Pro write long text in images?</h3><p>Yes — this is its strongest capability. It reproduced a four-line body paragraph, a headline, an italic subhead and a footer line in a single 2K image with every character correct, and at 4K it rendered a line of type roughly 12 pixels tall correctly. Quote the exact strings in your prompt to get the best results.</p><h3>Can I get a JPEG from Nano Banana Pro on Siray AI?</h3><p>No. The endpoint has no output_format parameter and every output we generated was PNG. A 4K image can be 19–26 MB, so plan storage accordingly or convert after download.</p><h3>Does Nano Banana Pro support image editing?</h3><p>Yes, through google/nano-banana-pro-i2i with 1 to 14 input images as URLs or data URLs, plus a prompt. Editing, compositing and style transfer from multiple references all work, and we tested the 14-image maximum successfully.</p><h3>Try Nano Banana Pro on Siray AI</h3><p>Create a free account, get an API key and run the typography poster prompt above — under ten cents for an image with a full paragraph of correct type. Open the <a href="https://console.siray.ai/model-detail/google/nano-banana-pro-t2i?ref=blog.siray.ai">Nano Banana Pro model page on Siray AI</a> to start.</p><h3>Sources</h3><ul><li><a href="https://blog.google/innovation-and-ai/products/nano-banana-pro/?ref=blog.siray.ai">Introducing Nano Banana Pro</a>, Google blog, published November 20, 2025, accessed November 12, 2026.</li><li><a href="https://deepmind.google/models/gemini-image/pro/?ref=blog.siray.ai">Gemini 3 Pro Image — Nano Banana Pro</a>, Google DeepMind model page, accessed November 12, 2026.</li><li><a href="https://docs.siray.ai/api-reference/model-api/nano-banana-pro-t2i?ref=blog.siray.ai">Nano Banana Pro t2i API reference</a>, Siray AI docs, accessed November 12, 2026.</li><li><a href="https://docs.siray.ai/api-reference/model-api/nano-banana-pro-i2i?ref=blog.siray.ai">Nano Banana Pro i2i API reference</a>, Siray AI docs, accessed November 12, 2026.</li><li>Siray AI model pricing API (GET /v1/models/pricing/{model}), queried November 12, 2026.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0420ee5b19c3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[MOSS Image API on Siray AI: Pricing, Features and Prompt Examples]]></title>
            <link>https://medium.com/@siary_ai/moss-image-api-on-siray-ai-pricing-features-and-prompt-examples-ebf0d8032aa3?source=rss-81e778058314------2</link>
            <guid isPermaLink="false">https://medium.com/p/ebf0d8032aa3</guid>
            <category><![CDATA[image-gen]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[t2i]]></category>
            <dc:creator><![CDATA[Siray AI]]></dc:creator>
            <pubDate>Wed, 18 Nov 2026 03:23:52 GMT</pubDate>
            <atom:updated>2026-11-18T03:23:52.331Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Xk1WJF0XqxHviy_Z.png" /></figure><p>The MOSS Image API gives developers OpenMOSS’s open-source text-to-image model through a standard REST call. On Siray AI it costs <strong>$0.02 to $0.08 per image</strong>, depending on size. It suits posters, infographics and other images that need readable English or Chinese text and controlled layout. This guide covers pricing, a quickstart and four tested prompts.</p><h3>Key takeaways</h3><ul><li><strong>One endpoint on Siray AI:</strong> text-to-image only, moss/moss-image-t2i. The model has no reference-image or editing input.</li><li><strong>Three price tiers:</strong> $0.02 at 1k, $0.04 at 1.5k and $0.08 at 2k.</li><li><strong>Quick turnaround in our tests:</strong> six tasks took 41 to 72 seconds, and each returned one image.</li><li><strong>Bilingual text held up:</strong> every English and Chinese line on our eight-line menu was correct.</li><li><strong>JSON layouts work through the </strong><strong>prompt field:</strong> all three elements of our test poster landed where we placed them.</li></ul><h3>What is MOSS Image?</h3><p>MOSS Image is an open-source text-to-image model family from the OpenMOSS team, <a href="https://www.moss.cn/en/models/moss-image">released on October 16, 2026</a>. It is OpenMOSS’s model, available on Siray AI through a unified API. The weights are released under Apache-2.0.</p><p>OpenMOSS describes these core capabilities:</p><ul><li><strong>Text rendering.</strong> Chinese and English text can appear as part of a scene, including headings, paragraphs, handwritten notes and board writing.</li><li><strong>Layout control.</strong> A JSON prompt can define objects, visible text and their bounding boxes.</li><li><strong>Long prompts.</strong> The <a href="https://github.com/OpenMOSS/MOSS-Image">model family accepts prompts of up to 4K tokens</a>.</li><li><strong>Varied styles.</strong> OpenMOSS shows anime, photorealistic images, sketches and graphic design.</li><li><strong>A compact backbone.</strong> The family is built on a 5B-parameter model, with Base and SFT variants. SFT is fine-tuned from Base on curated data.</li></ul><p>OpenMOSS also reports its own benchmark results. On LongLongTextBench, a benchmark its <a href="https://arxiv.org/abs/2610.23374">technical report</a> introduces, MOSS Image reached 96.3% mean text coverage, ranking first among the open-weight models evaluated. Coverage was 97.1% for the longest English texts and 92.4% for the longest Chinese texts. In end-to-end design comparisons on RealLayoutArena, its system ranked second. These are the authors’ figures, and we did not reproduce them.</p><p>The project is also clear about its limits:</p><ul><li><strong>No image input.</strong> The technical report says the backbone “does not directly condition on reference images or support image editing.”</li><li><strong>Prompt quality matters.</strong> The report’s prompt enhancer depends on external language models, whose errors can reduce quality.</li><li><strong>Layouts are experimental.</strong> The <a href="https://github.com/OpenMOSS/MOSS-Image/blob/main/USERGUIDE.md">official user guide</a> labels layout prompts “Experimental.”</li></ul><h3>MOSS Image endpoint on Siray AI</h3><p>Siray AI offers one MOSS Image endpoint:</p><ul><li><strong>moss/moss-image-t2i</strong> — Task: Text-to-image; Required inputs: prompt, size, aspect_ratio</li></ul><p>It accepts these values:</p><ul><li><strong>size</strong> — Values: 1k, 1.5k, 2k</li><li><strong>aspect_ratio</strong> — Values: 1:1, 1:2, 2:1, 1:3, 3:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 9:21, 21:9</li><li><strong>output_format (optional)</strong> — Values: jpg, png, webp</li></ul><p>The <a href="https://docs.siray.ai/api-reference/model-api/moss-image-t2i">Siray AI reference</a> lists no fields for negative prompts, seeds, steps or guidance. It also does not say whether the endpoint runs the Base or SFT variant, so do not assume either.</p><h3>MOSS Image pricing on Siray AI</h3><p>The price depends on the size you request:</p><ul><li><strong>1k</strong> — Listed price per image: $0.02</li><li><strong>1.5k</strong> — Listed price per image: $0.04</li><li><strong>2k</strong> — Listed price per image: $0.08</li></ul><p>Two worked examples:</p><ul><li><strong>100 drafts at </strong><strong>1k:</strong> 100 × $0.02 = <strong>$2.00</strong>.</li><li><strong>20 finals at </strong><strong>2k:</strong> 20 × $0.08 = <strong>$1.60</strong>.</li></ul><p>Prices were read from the Siray AI pricing API on November 5, 2026, with no discount active. Our test costs below use these listed prices. Check the <a href="https://console.siray.ai/model-detail/moss/moss-image-t2i?ref=blog.siray.ai">MOSS Image model page</a> for live prices. For broader context, see our <a href="https://blog.siray.ai/cheap-ai-video-image-api/">AI image and video API cost breakdown</a>.</p><h3>Quickstart: call the MOSS Image API</h3><p>Generation is asynchronous. You submit a task, get a task_id, then poll every 3–5 seconds until the status is SUCCESS. The image URL is in data.outputs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*60rwmtdqiphOBalh.png" /></figure><p><em>Illustration generated with MOSS Image on Siray AI.</em></p><p><strong>1. Submit a task</strong></p><pre>curl -X POST https://api.siray.ai/v1/images/generations/async \<br>  -H &quot;Authorization: Bearer $SIRAY_API_KEY&quot; \<br>  -H &quot;Content-Type: application/json&quot; \<br>  -d &#39;{<br>    &quot;model&quot;: &quot;moss/moss-image-t2i&quot;,<br>    &quot;prompt&quot;: &quot;Anime background art of a small seaside railway station at dusk, with no people. A single-track platform with a weathered wooden bench and a round clock runs along a calm ocean. A two-car local train waits with its windows glowing warm yellow, overhead wires cut across a violet and peach sky, and a white station sign on two posts reads exactly \&quot;SHIOMI\&quot;. Hand-painted gouache texture, soft cel shading, detailed clouds, and gentle reflections on the wet platform.&quot;,<br>    &quot;size&quot;: &quot;1k&quot;,<br>    &quot;aspect_ratio&quot;: &quot;3:2&quot;,<br>    &quot;output_format&quot;: &quot;png&quot;<br>  }&#39;<br>​</pre><p><strong>2. Poll for the result</strong></p><pre>curl https://api.siray.ai/v1/images/generations/async/$TASK_ID \<br>  -H &quot;Authorization: Bearer $SIRAY_API_KEY&quot;<br>​</pre><p>Two things we noticed while integrating:</p><ul><li><strong>size is a tier, not a pixel count.</strong> The output depends on the ratio, so read the real dimensions from the file.</li><li><strong>Do not time out early.</strong> Progress reached about 73–87%, then jumped to SUCCESS. Our slowest task took 72 seconds.</li></ul><p>Output sizes from our six tasks:</p><ul><li><strong>1k</strong> — aspect_ratio: 1:1; Output: 1024×1024</li><li><strong>1k</strong> — aspect_ratio: 3:2; Output: 1248×832</li><li><strong>1.5k</strong> — aspect_ratio: 3:4; Output: 1328×1760</li><li><strong>1.5k</strong> — aspect_ratio: 16:9; Output: 2048×1152</li><li><strong>2k</strong> — aspect_ratio: 16:9; Output: 2720×1536</li></ul><p>Status details are in the <a href="https://docs.siray.ai/api-reference/task-status-image">image task status API</a>.</p><h3>MOSS Image prompt examples (tested)</h3><p>We ran every prompt below through the Siray AI API on November 5, 2026, one at a time. Timings include queueing and were measured by a client polling every 5 seconds. Each result is the single unedited output from its task.</p><h3>1. Bilingual menu with exact text</h3><pre>A photograph of a black chalkboard menu hanging on the white-painted brick wall of a small neighborhood cafe, soft morning window light from the left, a potted fern on a wooden shelf below the board. The menu is hand-lettered in white chalk, with neatly aligned rows, and reads exactly, line by line:<br>&quot;MORNING MENU<br>早餐菜单<br>Oat Milk Latte  燕麦拿铁  $5.50<br>Sesame Toast  芝麻吐司  $4.00<br>Soft Boiled Eggs  溏心蛋  $3.50<br>Jasmine Cold Brew  茉莉冷萃  $6.00<br>Open daily 7:30 to 15:00<br>每天营业 7:30 至 15:00&quot;<br>Every word is sharp and legible, the English and Chinese are on the same row, and the prices line up in a right-hand column.<br>​</pre><p>Settings: size 1.5k, aspect_ratio 3:4, 41 seconds, $0.04, 1328×1760.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*bm4IzZB1WbWBOKWt.png" /></figure><p>Every line, price and time is correct, and the rows and price column line up. The model added chalk underlines and row rules we did not request, and the “O” in “MORNING” is smaller than its neighbors. This is the strongest result of our tests.</p><h3>2. Poster from a JSON layout</h3><pre>{&quot;summary&quot;:{&quot;meta&quot;:[&quot;poster&quot;],&quot;canvas&quot;:&quot;single&quot;,&quot;caption&quot;:&quot;A minimalist botanical exhibition poster with a warm cream background, a flat deep-green fern illustration in the center, a large serif title at the top, and a small date line at the bottom.&quot;},&quot;items&quot;:[{&quot;label&quot;:&quot;text&quot;,&quot;caption&quot;:&quot;Large deep-green serif uppercase title centered at the top.&quot;,&quot;text&quot;:&quot;QUIET GREEN&quot;,&quot;bbox&quot;:[-432,-384,-304,384]},{&quot;label&quot;:&quot;object&quot;,&quot;caption&quot;:&quot;A flat deep-green fern illustration with a single curved frond, centered.&quot;,&quot;bbox&quot;:[-224,-256,272,256]},{&quot;label&quot;:&quot;text&quot;,&quot;caption&quot;:&quot;Small deep-green sans-serif date line centered near the bottom.&quot;,&quot;text&quot;:&quot;Botanical Prints, 12 to 28 March&quot;,&quot;bbox&quot;:[352,-320,416,320]}],&quot;bbox_coord_space&quot;:&quot;abs_centralized&quot;,&quot;bbox_format&quot;:&quot;yxyx&quot;}<br>​</pre><p>Settings: size 1k, aspect_ratio 1:1, 41 seconds, $0.02, 1024×1024. Pass the whole JSON object as the prompt string.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zbPaLgOnFZFIphXv.png" /></figure><p>By visual check, the title, fern and date line all sit very close to their boxes, and both text strings are correct. One miss: the date line came out in a serif face, although its caption asked for sans-serif. This is a single simple layout, so test denser designs yourself.</p><h3>3. Photorealistic person and hands</h3><pre>A photorealistic wide shot of a ceramic workshop on a rainy afternoon. A potter in her sixties with short gray hair and a clay-stained canvas apron sits at a wheel, shaping a tall vase with both hands. Wooden shelves behind her hold rows of unglazed bowls. Soft overcast daylight comes through a large steel-framed window streaked with rain, leaving gentle shadows and a cool, muted palette with warm terracotta accents. Shot on a 35mm lens at f/2.8, shallow depth of field, natural skin texture, realistic wet clay on her fingers.<br>​</pre><p>Settings: size 2k, aspect_ratio 16:9, 57 seconds, $0.08, 2720×1536.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*t-lvv_MsammpaPcV.png" /></figure><p>The face, wet clay and rainy window follow the prompt well, and the visible hand looks natural. The shelves have dark metal uprights instead of being fully wooden, and the foreground buckets and back wall look smeared at full size. Review backgrounds before using this at large sizes.</p><h3>4. Anime scene with a short sign</h3><pre>Anime background art of a small seaside railway station at dusk, with no people. A single-track platform with a weathered wooden bench and a round clock runs along a calm ocean. A two-car local train waits with its windows glowing warm yellow, overhead wires cut across a violet and peach sky, and a white station sign on two posts reads exactly &quot;SHIOMI&quot;. Hand-painted gouache texture, soft cel shading, detailed clouds, and gentle reflections on the wet platform.<br>​</pre><p>Settings: size 1k, aspect_ratio 3:2, 41 seconds, $0.02, 1248×832.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ZzqMpIpbHPPHKntq.png" /></figure><p>The sign is spelled correctly, and the palette and painted look match the prompt. But silhouetted passengers appear in the train windows, despite “with no people.” The endpoint has no negative prompt field, so check exclusions like this one by eye.</p><h3>Prompting tips for MOSS Image</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*JmUlaOV8-INkDVv4.png" /></figure><p><em>Illustration generated with MOSS Image on Siray AI.</em></p><ol><li><strong>Set line breaks yourself.</strong> The official guide says the model can break lines automatically, “but the result may not always match your expectations.” Our menu used real line breaks inside quotes and came out correct.</li><li><strong>Describe the scene in English.</strong> OpenMOSS recommends English because it has the most training data. Our menu used an English scene description with Chinese visible text.</li><li><strong>Match layout coordinates to the output.</strong> abs_centralized measures pixels from the canvas center, and yxyx means [y_min, x_min, y_max, x_max]. 1k at 1:1 gave us 1024×1024, so each axis ran from −512 to +512.</li><li><strong>Be specific in long prompts.</strong> The technical report found that detailed content and design instructions improved human preference over concise requests.</li><li><strong>Do not rely on exclusions.</strong> “With no people” failed in our anime test. Describe a composition that leaves no room for the thing you want to avoid.</li><li><strong>Draft at </strong><strong>1k, then scale up.</strong> A 2k image costs four times as much as 1k. There is no seed field in the docs, so a re-run at a larger size may not match your draft.</li></ol><h3>When to use MOSS Image (and when not to)</h3><p><strong>Good fit:</strong></p><ul><li>Posters, menus and signs with exact English or Chinese text.</li><li>Graphics where elements must sit in planned positions.</li><li>Anime and photographic concepts, reviewed by a person.</li><li>High-volume drafts at $0.02 per 1k image.</li></ul><p><strong>Consider another model when:</strong></p><ul><li>You need to change an existing image. MOSS Image has no image input, so see the <a href="https://blog.siray.ai/qwen-image-3-edit-api-prompt-guide/">Qwen Image 3 Edit API guide</a>.</li><li>You need text on real packaging. See <a href="https://blog.siray.ai/gpt-image-2-5-flare-packaging-mockup/">GPT Image 2.5 Flare for packaging localization</a>.</li><li>You want a specific print style with composition checks. See <a href="https://blog.siray.ai/flux-kontext-pro-woodcut-posters/">FLUX Kontext Pro woodcut posters</a>.</li><li>You need negative prompts or fixed seeds. The Siray AI endpoint does not list them.</li></ul><h3>FAQ</h3><h3>How much does the MOSS Image API cost per image?</h3><p>Siray AI lists MOSS Image at $0.02 for 1k, $0.04 for 1.5k and $0.08 for 2k. Prices were checked on November 5, 2026, with no discount active.</p><h3>Is MOSS Image available via API?</h3><p>Yes. On Siray AI, use moss/moss-image-t2i through the async image endpoint with one API key.</p><h3>Does MOSS Image support image editing or reference images?</h3><p>No. The technical report says the current backbone does not condition on reference images or support image editing, and the Siray AI endpoint is text-to-image only.</p><h3>Can MOSS Image render Chinese and English text?</h3><p>Yes. OpenMOSS lists both, and our eight-line bilingual menu came out correct. OpenMOSS’s own benchmark shows lower coverage for its longest Chinese texts (92.4%) than for its longest English texts (97.1%).</p><h3>How do I use JSON layouts with the MOSS Image API?</h3><p>Pass the layout JSON as the prompt string. Use the official summary and items format, with bounding boxes that match the output dimensions. OpenMOSS labels this feature experimental.</p><h3>How long does MOSS Image take to generate an image?</h3><p>Our six tasks took 41 to 72 seconds, including queue time. The 2k task took 57 seconds, and the 1k tasks took 41 seconds.</p><h3>Try MOSS Image on Siray AI</h3><p>Create an account, get an API key and run the chalkboard menu prompt above. Open the <a href="https://console.siray.ai/model-detail/moss/moss-image-t2i?ref=blog.siray.ai">MOSS Image model page on Siray AI</a> to start.</p><h3>Sources</h3><ul><li><a href="https://www.moss.cn/en/models/moss-image">MOSS-Image model page</a>, MOSS, accessed November 5, 2026.</li><li><a href="https://github.com/OpenMOSS/MOSS-Image">OpenMOSS/MOSS-Image repository</a>, GitHub, accessed November 5, 2026.</li><li><a href="https://github.com/OpenMOSS/MOSS-Image/blob/main/USERGUIDE.md">MOSS-Image user guide</a>, OpenMOSS, accessed November 5, 2026.</li><li><a href="https://arxiv.org/abs/2610.23374">MOSS-Image Technical Report</a>, arXiv, accessed November 5, 2026.</li><li><a href="https://docs.siray.ai/api-reference/model-api/moss-image-t2i">Moss Image t2i API reference</a>, Siray AI docs, accessed November 5, 2026.</li><li><a href="https://docs.siray.ai/api-reference/task-status-image">Image task status API</a>, Siray AI docs, accessed November 5, 2026.</li><li>Siray AI model pricing API (GET /v1/models/pricing/{model}), queried November 5, 2026.</li></ul><p><em>Originally published on the </em><a href="https://blog.siray.ai/moss-image-api-pricing-prompts/"><em>Siray AI blog</em></a><em>. Prices were checked on 2026–11–05 and can change; see the live model page for current pricing.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ebf0d8032aa3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Seedance 2.5 Flash Spicy API on Siray AI: Pricing, Endpoints and Tested Prompts]]></title>
            <link>https://medium.com/@siary_ai/seedance-2-5-flash-spicy-api-on-siray-ai-pricing-endpoints-and-tested-prompts-ef304c6b25a4?source=rss-81e778058314------2</link>
            <guid isPermaLink="false">https://medium.com/p/ef304c6b25a4</guid>
            <dc:creator><![CDATA[Siray AI]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 02:54:49 GMT</pubDate>
            <atom:updated>2026-11-17T03:23:18.997Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="Seedance 2.5 Flash Spicy API on Siray AI: Pricing, Endpoints and Tested Prompts" src="https://cdn-images-1.medium.com/max/1024/1*w5jTIrADqmOee4OfX_6MfA.png" /><figcaption>Seedance 2.5 Flash Spicy API on Siray AI: Pricing, Endpoints and Tested Prompts</figcaption></figure><p>The Seedance 2.5 Flash Spicy API gives you ByteDance&#39;s Seedance 2.5 Flash video model through Siray AI endpoints intended for mature creative work. It costs <strong>about $0.068 per second at 480p and $0.151 per second at 720p</strong>, for clips of 4 to 30 seconds with optional audio. It suits fast drafts, high-volume social clips and adult-audience projects.</p><h3>Key takeaways</h3><ul><li><strong>Three Spicy endpoints:</strong> text-to-video, image-to-video (first frame, optional last frame) and reference-to-video with up to 10 images, 5 videos and 5 audio files.</li><li><strong>Token-based pricing:</strong> $0.007 per 1K tokens. A 5-second 720p clip costs about $0.76 and a 30-second 720p clip about $4.54.</li><li><strong>Fast in our tests:</strong> seven clips of 5 to 10 seconds each finished in 72 to 100 seconds from submit to result, including five submitted at the same time.</li><li><strong>480p and 720p only:</strong> there is no 1080p tier. Use the full Seedance 2.5 when you need higher resolution.</li><li><strong>Output moderation still applies:</strong> one of our SFW prompts (a neon city street) was rejected with a copyright-related policy error, and that task was not billed.</li></ul><h3>What is Seedance 2.5 Flash?</h3><p>Seedance 2.5 Flash is a faster, cheaper tier of ByteDance&#39;s Seedance 2.5 video model. ByteDance&#39;s cloud unit Volcano Engine announced it on November 12, 2026, and API access opened on November 15, 2026, according to <a href="https://pandaily.com/bytedance-seedance-2-5-flash-faster-lower-cost-video-model-volcano-ark-api">Pandaily</a> and other launch coverage. ByteDance says Flash generates video about twice as fast as Seedance 2.5 at roughly 34% lower cost, with output close to the flagship. We could not open ByteDance&#39;s own Flash pages to confirm these figures, so treat them as reported vendor claims.</p><p>The flagship <a href="https://seed.bytedance.com/en/seedance2_5">Seedance 2.5</a> is described by ByteDance Seed as &quot;a next-generation audio-video joint generation model, built for 30-second storytelling&quot;, with up to 30 seconds per generation, more reliable editing and better understanding of reference videos. Flash keeps the 30-second length and native audio but, per launch coverage, drops to 720p and accepts fewer references.</p><p><strong>Known limits.</strong> Flash outputs only 480p and 720p. Siray AI&#39;s reference endpoint accepts up to 10 images, 5 videos and 5 audio files, fewer than the flagship&#39;s reported 50 references. ByteDance&#39;s own speed claims have not been independently benchmarked; our timings below are one data point.</p><h3>What does Spicy mean?</h3><p>Spicy is a Siray AI catalog label, not a separate ByteDance model. The <a href="https://docs.siray.ai/api-reference/model-api/seedance-2.5-flash-t2v-spicy">Siray AI API reference</a> says the variant &quot;supports mature and boundary-pushing creative content for production workflows.&quot; On Siray AI, Seedance 2.5 Flash is currently available only as Spicy endpoints.</p><p>Fewer restrictions do not mean no rules. Siray AI&#39;s <a href="https://blog.siray.ai/uncensored-ai-generation/">policy for uncensored generation</a> says all content &quot;must depict legal adult (18+) individuals only&quot;, and illegal content is refused. Don&#39;t use a real person&#39;s likeness without their consent, and follow your local laws and the rules of the platforms where you publish. Every example in this guide is SFW. For an adult-oriented walkthrough, see <a href="https://blog.siray.ai/seedance-2-5-spicy-slow-dance/">Seedance 2.5 Spicy Slow Dance: Direct Two Adult Partners</a>.</p><h3>Seedance 2.5 Flash Spicy endpoints on Siray AI</h3><p>Model IDTaskRequired inputsOptional inputsbytedance/seedance-2.5-flash-t2v-spicyText-to-videoprompt, duration, size, aspect_ratioaudio_enablebytedance/seedance-2.5-flash-i2v-spicyImage-to-videoprompt, image, duration, size, aspect_ratioend_image, audio_enablebytedance/seedance-2.5-flash-ref2v-spicyReference-to-video, edit, extendprompt, duration, size, aspect_ratioimages (≤10), videos (≤5), audios (≤5), task_type, audio_enable</p><p>Shared values on all three endpoints:</p><ul><li>duration: any whole number from 4 to 30. The reference endpoint also accepts -1, which lets the model choose the length.</li><li>size: 480p or 720p.</li><li>aspect_ratio: adaptive, 16:9, 4:3, 1:1, 3:4, 9:16 or 21:9.</li><li>task_type (reference endpoint only): reference, auto, edit or extend.</li><li>Images can be URLs or data URLs.</li></ul><h3>Seedance 2.5 Flash Spicy pricing on Siray AI</h3><p>All three endpoints are billed by output tokens at <strong>$0.007 per 1K tokens</strong>. Siray AI&#39;s pricing note gives the formula: height × width × (input duration + output duration) × 24 ÷ 1024. In practice that works out to these prices for text-to-video, image-to-video and reference-to-video without video input:</p><p>Duration480p720p5 seconds$0.34$0.75610 seconds$0.68$1.51215 seconds$1.02$2.26830 seconds$2.04$4.536Per second≈ $0.068≈ $0.151</p><p>When you pass reference videos, the rate drops to $0.0043 per 1K tokens, but the input video&#39;s length is added to the token count. Siray AI lists a range for these jobs; for example, a 5-second 720p reference job with video input costs $0.84 to $3.25 depending on the input.</p><p>Worked examples:</p><ul><li><strong>A 5-second 720p draft:</strong> $0.756. Twenty drafts cost $15.12.</li><li><strong>A 15-second 9:16 social clip at 480p:</strong> $1.02.</li><li><strong>A full 30-second 720p spot:</strong> $4.536.</li></ul><p>Prices were read from the Siray AI pricing API on November 16, 2026, with no discount active. Check the <a href="https://console.siray.ai/model-detail/bytedance/seedance-2.5-flash-t2v-spicy?ref=blog.siray.ai">Seedance 2.5 Flash t2v Spicy model page</a> for live prices. In our tests, actual charges matched the formula within about 1%: a 5-second 720p clip used 108,900 tokens ($0.762).</p><p><em>Concept video generated with Seedance 2.5 Flash on Siray AI.</em></p><h3>Quickstart: call the Seedance 2.5 Flash Spicy API</h3><p>Generation is asynchronous. Submit a task to get a task_id, then check it every 3–5 seconds. When the status is SUCCESS, download the video from data.outputs.</p><p><strong>1. Submit a task</strong></p><pre>curl -X POST https://api.siray.ai/v1/video/generations \<br>  -H &quot;Authorization: Bearer $SIRAY_API_KEY&quot; \<br>  -H &quot;Content-Type: application/json&quot; \<br>  -d &#39;{<br>    &quot;model&quot;: &quot;bytedance/seedance-2.5-flash-t2v-spicy&quot;,<br>    &quot;prompt&quot;: &quot;A slow cinematic drone push-in over a foggy pine forest at sunrise...&quot;,<br>    &quot;duration&quot;: 5,<br>    &quot;size&quot;: &quot;720p&quot;,<br>    &quot;aspect_ratio&quot;: &quot;16:9&quot;,<br>    &quot;audio_enable&quot;: true<br>  }&#39;<br>​</pre><p>The response contains the task ID. In our tests it was at the top level of the response (task_id), not under data as the API reference shows, so read both places.</p><p><strong>2. Poll until the task finishes</strong></p><pre>curl https://api.siray.ai/v1/video/generations/&lt;task_id&gt; \<br>  -H &quot;Authorization: Bearer $SIRAY_API_KEY&quot;<br>​</pre><p>status goes from IN_PROGRESS to SUCCESS or FAILURE. On success, data.outputs holds the MP4 URL and data.usage.completion_tokens shows the billed tokens. On failure, check fail_code and fail_reason.</p><p>For image-to-video, change model to bytedance/seedance-2.5-flash-i2v-spicy and add &quot;image&quot;: &quot;&lt;URL or data URL&gt;&quot;. For reference-to-video, use bytedance/seedance-2.5-flash-ref2v-spicy with &quot;task_type&quot;: &quot;reference&quot; and &quot;images&quot;: [...].</p><p><em>Concept video generated with Seedance 2.5 Flash on Siray AI.</em></p><h3>Seedance 2.5 Flash prompt examples (tested)</h3><p>We ran every example below on November 16, 2026 through the Siray AI API. Times are measured from submit to finish on the task record. Costs come from the billed tokens.</p><h3>1. Drone camera move with ambient audio</h3><pre>A slow cinematic drone push-in over a foggy pine forest at sunrise. The camera starts high above the treetops, glides forward and gradually tilts down to reveal a narrow river winding through the valley, golden light breaking through the mist. Smooth, stable motion, no cuts. Natural ambient sound: soft wind, distant birdsong, the faint rush of water. No people.<br>​</pre><ul><li>Endpoint: t2v-spicy, 5 s, 720p, 16:9, audio_enable: true</li><li>Time: 99 seconds · Cost: $0.762 · Output: 1280×720, 24 fps, with an audio track</li></ul><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1247479295%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1247479295%3Ffl%3Dpl%26fe%3Dsh&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2227265030-dae31b16f50cb8900275a708a9edada72e6fa025a501ea4cc9f205ee65e8b2bb-d_1280%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="1280" height="720" frameborder="0" scrolling="no"><a href="https://medium.com/media/65a36d9708fa6b775a0c9572d38911a9/href">https://medium.com/media/65a36d9708fa6b775a0c9572d38911a9/href</a></iframe><p>The push-in and the tilt down to the river came out exactly as written, with no jitter. Spelling out the start point, the motion and the reveal kept the camera move readable in five seconds.</p><h3>2. Character action at 480p</h3><pre>Medium shot in a sunlit pottery studio. A woman in her early 30s with a linen apron and rolled-up sleeves centers wet clay on a spinning wheel, presses her thumbs into the middle and slowly pulls the walls of a bowl upward. Her hands stay on the clay the whole time; the bowl rises steadily as the wheel turns. Static camera at table height, shallow depth of field, warm afternoon light through a dusty window. Adults only, no children, teenagers, babies or minors.<br>​</pre><ul><li>Endpoint: t2v-spicy, 8 s, 480p, 16:9, no audio</li><li>Time: 85 seconds · Cost: $0.541 · Output: 854×480, 24 fps</li></ul><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1247479323%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1247479323&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2227265026-8ede36c8bce80bc78cf6d5aa2a3f130432d1717db83f3ba7491e6d7c3fa31c75-d_640%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/bad0f2b5d74173ddd325df35fa349ae9/href">https://medium.com/media/bad0f2b5d74173ddd325df35fa349ae9/href</a></iframe><p>Her hands stayed on the clay for all eight seconds, which is where many video models fail. The clay grew into a tall vase rather than a bowl, so name the final shape more firmly if it matters. 480p is clearly softer, but it is enough to check motion before paying for 720p.</p><h3>3. Three-shot sequence with sound</h3><pre>A three-shot sequence at a small seaside bakery at dawn, natural documentary look.<br>Shot 1 (wide): a baker in her 50s with grey hair tied back and a flour-dusted blue apron unlocks the wooden door of a whitewashed bakery by the harbor and steps inside.<br>Shot 2 (close-up): her hands shape a round loaf of dough on a floured wooden table, then score the top with a small blade.<br>Shot 3 (medium shot): she slides the loaf into a glowing brick oven with a long wooden peel and smiles as warm light falls on her face.<br>Keep the same woman, apron and bakery in every shot. Sound: seagulls outside, the creak of the door, the crackle of the oven fire. Adults only, no children, teenagers, babies or minors. No text, no logos.<br>​</pre><ul><li>Endpoint: t2v-spicy, 10 s, 720p, 16:9, audio_enable: true</li><li>Time: 100 seconds · Cost: $1.518 · Output: 1280×720, with an audio track</li></ul><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1247479292%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1247479292%3Ffl%3Dtl%26fe%3Dec&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2227265037-d934855c2586a15f45f16e5e7a69b62eddb7d69d3ab13312257b9892f4e1c72d-d_1280%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="1280" height="720" frameborder="0" scrolling="no"><a href="https://medium.com/media/740e5f071a30376b454a946499328789/href">https://medium.com/media/740e5f071a30376b454a946499328789/href</a></iframe><p>All three shots appeared in order with clean cuts. The blue apron and grey hair matched across shots. Numbering shots and giving each a framing word (wide, close-up, medium) was enough to get the edit.</p><p>Our first version of this prompt, a man walking through a rainy neon-lit city, failed with OutputVideoSensitiveContentDetected.PolicyViolation and the message &quot;the output video may be related to copyright restrictions.&quot; The prompt was SFW. Our guess is that busy signage triggered the copyright check. No tokens were reported for the failed task.</p><h3>4. Image-to-video: animate a product photo</h3><pre>Animate this product photo. Steam begins to rise slowly from the mug and curls upward; soft morning light shifts gently across the table. The camera makes a slow, smooth orbit of about 20 degrees to the right around the mug. The mug itself stays still and keeps its exact shape, color and details. No hands, no people, no text.<br>​</pre><ul><li>Endpoint: i2v-spicy, 5 s, 720p, aspect_ratio: adaptive, one input image (data URL)</li><li>Time: 72 seconds · Cost: $0.766 · Output: 1178×786</li></ul><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1247479322%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1247479322%3Ffl%3Dtl%26fe%3Dec&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2227265065-1b8e495ccde23d701941d074672ba7c8183546bebe50bd7bbe9b74bdceeb2d39-d_960%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="1178" height="786" frameborder="0" scrolling="no"><a href="https://medium.com/media/862d48693b39d8e331f1b81d462539e9/href">https://medium.com/media/862d48693b39d8e331f1b81d462539e9/href</a></iframe><p>The mug kept its shape, glaze and handle exactly, and the steam looked natural. The orbit was much smaller than 20 degrees. With adaptive, the output followed the input photo&#39;s ratio (1178×786) instead of a fixed 1280×720.</p><h3>5. Reference-to-video: place a product in a location</h3><pre>Use image 1 as the location and image 2 as the product. A slow dolly shot across the bright salt flat from image 1 at golden hour; the ceramic mug from image 2 sits on a small wooden stool in the middle of the salt flat, keeping its exact shape and color. Warm wind moves thin clouds across the sky and the long shadow of the stool stretches toward the camera. Calm, minimal, editorial mood. No people, no text.<br>​</pre><ul><li>Endpoint: ref2v-spicy, 6 s, 720p, 16:9, task_type: reference, two input images</li><li>Time: 84 seconds · Cost: $0.914 · Output: 1280×720</li></ul><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1247479356%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1247479356%3Ffl%3Dtl%26fe%3Dec&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2227265072-95853bfdf4db4bf37abbb8c5cd45fd1861446fec038aa0cc2b7215b58aa98aa1-d_960%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="1280" height="720" frameborder="0" scrolling="no"><a href="https://medium.com/media/1c01c95461501871cbabccd19122e1e1/href">https://medium.com/media/1c01c95461501871cbabccd19122e1e1/href</a></iframe><p>Referring to inputs as &quot;image 1&quot; and &quot;image 2&quot; worked: the model took the location from one image and the mug from the other. The mug&#39;s colour and handle matched the reference. Camera movement was again gentler than asked.</p><h3>Prompting tips for Seedance 2.5 Flash</h3><ul><li><strong>Ask for more camera motion than you want.</strong> In both image-driven tests, orbits and dollies came out smaller than written. Pure text-to-video followed camera directions closely.</li><li><strong>Number your shots.</strong> &quot;Shot 1 (wide) … Shot 2 (close-up) …&quot; produced clean cuts in a 10-second clip. Repeat the details that must stay the same in every shot.</li><li><strong>Name inputs by position.</strong> On the reference endpoint, &quot;image 1 as the location, image 2 as the product&quot; was enough to assign roles.</li><li><strong>Draft at 480p.</strong> A 480p clip costs 45% of the 720p price and shows whether motion and timing work.</li><li><strong>Keep busy signage out of SFW prompts.</strong> Neon signs and dense brand-like text may trigger the copyright check, as in our failed test.</li><li><strong>State adult ages and add &quot;no children, teenagers, babies or minors&quot;.</strong> This is required by Siray AI&#39;s rules for Spicy endpoints and keeps results predictable.</li></ul><h3>When to use Seedance 2.5 Flash (and when not to)</h3><p><strong>Use it for:</strong></p><ul><li>Fast drafts and A/B variants before a final render.</li><li>High-volume social clips at 480p or 720p.</li><li>Product clips from a photo or from reference images.</li><li>Mature-themed creative work that needs a Spicy endpoint.</li></ul><p><strong>Look elsewhere when:</strong></p><ul><li>You need 1080p or higher: use the full Seedance 2.5 or another model that offers 1080p.</li><li>You need more than 10 reference images in one request.</li><li>You need a non-Spicy endpoint for compliance reasons: the plain bytedance/seedance-2.5-t2v, i2v and ref2v endpoints are also on Siray AI.</li></ul><p>For a reference-driven example with the full model, see <a href="https://blog.siray.ai/seedance-2-5-museum-clips-parallax-without-reshaping/">Seedance 2.5 Museum Clips: Parallax Without Reshaping</a>.</p><h3>FAQ</h3><h3>How much does the Seedance 2.5 Flash Spicy API cost per second?</h3><p>About $0.068 per second at 480p and $0.151 per second at 720p on Siray AI. Billing is $0.007 per 1K output tokens, so a 5-second 720p clip costs about $0.76 (prices checked November 16, 2026).</p><h3>What is the difference between Seedance 2.5 Flash and Seedance 2.5?</h3><p>Flash is ByteDance&#39;s faster, cheaper tier of Seedance 2.5. ByteDance says it is about twice as fast and about 34% cheaper, but it outputs only 480p or 720p and takes fewer reference files.</p><h3>How long can a Seedance 2.5 Flash video be?</h3><p>From 4 to 30 seconds, chosen with duration. The reference-to-video endpoint also accepts -1 to let the model pick the length.</p><h3>Does Seedance 2.5 Flash generate audio?</h3><p>Yes. Set audio_enable: true. In our tests, clips with audio cost the same as the token formula predicts for the video alone.</p><h3>Does Seedance 2.5 Flash support image-to-video?</h3><p>Yes. Use bytedance/seedance-2.5-flash-i2v-spicy with image as the first frame and optional end_image as the last frame. The reference endpoint accepts up to 10 images, 5 videos and 5 audio files.</p><h3>How long does Seedance 2.5 Flash take to generate a video?</h3><p>In our tests, 5- to 10-second clips took 72 to 100 seconds from submit to finish, and five tasks submitted at the same time did not slow each other down.</p><h3>Try Seedance 2.5 Flash Spicy on Siray AI</h3><p>Get an API key and run your first clip from the <a href="https://console.siray.ai/model-detail/bytedance/seedance-2.5-flash-t2v-spicy?ref=blog.siray.ai">Seedance 2.5 Flash t2v Spicy model page</a>. The same key works for the image-to-video and reference-to-video endpoints and for every other model on <a href="https://console.siray.ai/?ref=blog.siray.ai">Siray AI</a>.</p><h3>Sources</h3><ul><li>ByteDance Seed, <a href="https://seed.bytedance.com/en/seedance2_5">Seedance 2.5</a>, accessed November 16, 2026.</li><li>BytePlus, <a href="https://www.byteplus.com/en/blog/dreamina-seedance2-5">Dreamina Seedance 2.5 announcement</a>, accessed November 16, 2026.</li><li>Pandaily, <a href="https://pandaily.com/bytedance-seedance-2-5-flash-faster-lower-cost-video-model-volcano-ark-api">ByteDance Releases Seedance 2.5 Flash</a>, accessed November 16, 2026.</li><li>Siray AI API reference: <a href="https://docs.siray.ai/api-reference/model-api/seedance-2.5-flash-t2v-spicy">t2v Spicy</a>, <a href="https://docs.siray.ai/api-reference/model-api/seedance-2.5-flash-i2v-spicy">i2v Spicy</a>, <a href="https://docs.siray.ai/api-reference/model-api/seedance-2.5-flash-ref2v-spicy">ref2v Spicy</a>, accessed November 16, 2026.</li><li>Siray AI pricing API, GET /v1/models/pricing/{model}, November 16, 2026.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ef304c6b25a4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Midjourney V8.2 API on Siray AI: Pricing, Features and Prompt Examples]]></title>
            <link>https://medium.com/@siary_ai/midjourney-v8-2-api-on-siray-ai-pricing-features-and-prompt-examples-d0dcd1c4275d?source=rss-81e778058314------2</link>
            <guid isPermaLink="false">https://medium.com/p/d0dcd1c4275d</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[midjourney]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[api]]></category>
            <dc:creator><![CDATA[Siray AI]]></dc:creator>
            <pubDate>Wed, 04 Nov 2026 07:50:10 GMT</pubDate>
            <atom:updated>2026-11-04T07:50:10.982Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="midjourney-v8–2-api" src="https://cdn-images-1.medium.com/max/1000/1*q2NbgG8EOsTv60r1Ib4oRw.png" /><figcaption>midjourney-v8–2-api</figcaption></figure><h3>What is Midjourney V8.2?</h3><p>Midjourney V8.2 is a closed image model from Midjourney, <a href="https://updates.midjourney.com/version-8-2/">launched on July 24, 2026</a>. It is Midjourney’s model, available on Siray AI through a unified API.</p><p>Midjourney says the release “focuses on aesthetics and image quality and personalization.” Its announcement makes three specific claims:</p><ul><li><strong>Look.</strong> Images “should be more creative, bold, sophisticated, edgy and fresh.”</li><li><strong>Consistency.</strong> “Random situations where you’d get a low quality image should be dramatically reduced.”</li><li><strong>Personalization.</strong> The model should “understand” personal taste much better, especially with many ratings under a profile.</li></ul><p>Personalization is tied to Midjourney profiles, and the Siray AI parameter list has no field for it.</p><p>Midjourney also describes <a href="https://updates.midjourney.com/edit-model-for-v8/">a separate V8.2 edit model</a> with instruction-based editing, inpainting and outpainting, and up to four image references. It warned of “a lot of edge cases” during testing. The Siray AI docs do not say whether the i2i endpoint uses this edit model, so do not assume those features are available through it.</p><h3>Midjourney V8.2 endpoints on Siray AI</h3><ul><li><strong>midjourney/midjourney-v8.2-t2i</strong> — Task: Text-to-image; Required inputs: prompt, aspect_ratio</li><li><strong>midjourney/midjourney-v8.2-i2i</strong> — Task: Image-to-image; Required inputs: prompt, aspect_ratio, image (URL or data URL)</li></ul><p>Both endpoints share these optional parameters:</p><ul><li><strong>stylize</strong> — Values: 0–1000, default 100; What it does: Strength of Midjourney&#39;s aesthetic</li><li><strong>raw</strong> — Values: true / false; What it does: Reduces default aesthetic processing</li><li><strong>chaos</strong> — Values: 0–100; What it does: Varies the initial image grids</li><li><strong>weird</strong> — Values: 0–3000; What it does: Adds unusual aesthetics</li><li><strong>negative_prompt</strong> — Values: text; What it does: Lists content to exclude</li><li><strong>seed</strong> — Values: -1 to 9999999999; What it does: Sets the random seed</li><li><strong>style_reference</strong> — Values: URL, number or random; What it does: Provides a style reference</li><li><strong>style_weight</strong> — Values: 0–1000, default 100; What it does: Weight of the style reference</li><li><strong>quality</strong> — Values: 1, 2, 4; What it does: Sets the quality option</li><li><strong>tile</strong> — Values: true / false; What it does: Creates repeating tile images</li></ul><p>We tested raw, stylize and negative_prompt. The other parameters are listed from the docs and were not part of our tests.</p><h3>Midjourney V8.2 pricing on Siray AI</h3><ul><li><strong>Text-to-image</strong> — Listed price: About $0.021 per image</li><li><strong>Image-to-image</strong> — Listed price: Not returned by the pricing API at the time of writing</li></ul><p>The pricing description reads “billed at approximately $0.021 each,” and it does not say whether a four-image task counts as one image or four. Until you confirm this on your own usage page, budget for the higher figure:</p><ul><li><strong>One task:</strong> $0.021 if billed once, or 4 × $0.021 = <strong>$0.084</strong> if billed per output.</li><li><strong>100 tasks:</strong> $2.10 if billed once, or <strong>$8.40</strong> if billed per output.</li></ul><p>Prices were read from the Siray AI pricing API on November 4, 2026, with no discount active. Check the <a href="https://console.siray.ai/model-detail/midjourney/midjourney-v8.2-t2i?ref=blog.siray.ai">Midjourney V8.2 model page</a> for live prices. For broader context, see our <a href="https://blog.siray.ai/cheap-ai-video-image-api/">AI image and video API cost breakdown</a>.</p><h3>Quickstart: call the Midjourney V8.2 API</h3><p>Generation is asynchronous. You submit a task, get a task_id, then poll every 3–5 seconds until the status is SUCCESS. The image URLs are in data.outputs.</p><figure><img alt="Illustration generated with Midjourney V8.2 on Siray AI." src="https://cdn-images-1.medium.com/max/1024/1*vN2uGYr-1vHuaMps0lV_Qw.png" /><figcaption><em>Illustration generated with Midjourney V8.2 on Siray AI.</em></figcaption></figure><p><strong>1. Submit a task</strong></p><pre>curl -X POST https://api.siray.ai/v1/images/generations/async \<br>  -H &quot;Authorization: Bearer $SIRAY_API_KEY&quot; \<br>  -H &quot;Content-Type: application/json&quot; \<br>  -d &#39;{<br>    &quot;model&quot;: &quot;midjourney/midjourney-v8.2-t2i&quot;,<br>    &quot;prompt&quot;: &quot;Studio product photograph of a matte sage-green ceramic coffee mug on a pale travertine block, a thin curl of steam rising, soft window light from the left, gentle shadow, warm neutral background, shallow depth of field, crisp detail on the glaze texture, e-commerce hero shot.&quot;,<br>    &quot;aspect_ratio&quot;: &quot;3:2&quot;,<br>    &quot;raw&quot;: true<br>  }&#39;<br>​</pre><p><strong>2. Poll for the result</strong></p><pre>curl https://api.siray.ai/v1/images/generations/async/$TASK_ID \<br>  -H &quot;Authorization: Bearer $SIRAY_API_KEY&quot;<br>​</pre><p>Three things we noticed while integrating:</p><ul><li><strong>Do not resubmit at 98%.</strong> Progress stayed there for minutes in the runs we logged. Set your client timeout to at least 10 minutes.</li><li><strong>Check </strong><strong>status, not </strong><strong>fail_reason.</strong> On our successful tasks, fail_reason contained the output URLs instead of being empty as the docs example shows.</li><li><strong>Do not hard-code pixel sizes.</strong> Most of our 16:9 tasks returned 1456x816, but one returned 1024x573. Our 3:2 tasks returned 1344x896.</li></ul><p>Full details are in the <a href="https://docs.siray.ai/api-reference/model-api/midjourney-v8.2-t2i">Midjourney V8.2 t2i API reference</a> and the <a href="https://docs.siray.ai/api-reference/task-status-image">image task status API</a>.</p><h3>Midjourney V8.2 prompt examples (tested)</h3><p>We ran every prompt below through the Siray AI API on November 4, 2026. Timings include queueing. Each result is one unedited output from the task.</p><h3>1. Product photo with raw mode</h3><pre>Studio product photograph of a matte sage-green ceramic coffee mug on a pale travertine block, a thin curl of steam rising, soft window light from the left, gentle shadow, warm neutral background, shallow depth of field, crisp detail on the glaze texture, e-commerce hero shot.<br>​</pre><figure><img alt="Sage-green ceramic mug with rising steam on a travertine block in window light, generated by Midjourney V8.2" src="https://cdn-images-1.medium.com/max/1024/1*viifu8_Egd8Yis3u3XBnKQ.png" /><figcaption>Sage-green ceramic mug with rising steam on a travertine block in window light, generated by Midjourney V8.2</figcaption></figure><p>Settings: text-to-image, aspect_ratio 3:2, raw true, 306 seconds, four outputs at 1344x896.</p><p>The light direction, glaze and stone texture all follow the prompt. The mug’s shape changed between candidates, from a wide cup to a straight mug, so this works better for concepts than for matching a real product.</p><h3>2. Poster with short exact text</h3><pre>A minimalist Swiss-style concert poster on off-white paper. Large bold black sans-serif headline with the text &quot;NIGHT SIGNALS&quot; across the top, a single cobalt blue circle in the center, and smaller text at the bottom reading &quot;LIVE AT HALL 7&quot;. Clean grid layout, generous margins, subtle paper texture, photographed flat under soft daylight.<br>​</pre><figure><img alt="Swiss-style poster reading NIGHT SIGNALS and LIVE AT HALL 7 with a cobalt blue circle, generated by Midjourney V8.2" src="https://cdn-images-1.medium.com/max/1024/1*ARc56uXIs_khzvHyACXwlQ.png" /><figcaption>Swiss-style poster reading NIGHT SIGNALS and LIVE AT HALL 7 with a cobalt blue circle, generated by Midjourney V8.2</figcaption></figure><p>Settings: text-to-image, aspect_ratio 3:2, raw true, stylize 50, 192 seconds, four outputs at 1344x896.</p><p>All four candidates spelled both lines correctly and kept the layout. Letter spacing and letter shapes were uneven in some candidates, and some added fold lines we did not ask for. Check them at full size before use.</p><h3>3. Painterly image with high stylize</h3><pre>A lighthouse on a black basalt cliff during a storm at dusk, painted in thick impasto oil, sweeping beam of pale gold light cutting through rain, towering teal waves, dramatic and dreamlike, gallery painting.<br>​</pre><figure><img alt="Lighthouse on a dark cliff above stormy teal waves in thick oil paint, generated by Midjourney V8.2" src="https://cdn-images-1.medium.com/max/1024/1*-ihgZBedVw__HlCtfEmXWQ.png" /><figcaption>Lighthouse on a dark cliff above stormy teal waves in thick oil paint, generated by Midjourney V8.2</figcaption></figure><p>Settings: text-to-image, aspect_ratio 3:2, stylize 750, 313 seconds, four outputs at 1344x896.</p><p>This is the strongest result of our tests. The brushwork, the palette and the beam all match, and the candidates differed in composition, which gives you real options from a single task.</p><h3>4. Image-to-image scene change</h3><pre>The same mug and travertine block, now outdoors on a sunlit stone garden wall at golden hour, soft olive leaves blurred in the background, warm evening light from the right.<br>​</pre><figure><img alt="Two-by-two grid of near-identical sage-green mugs still in the original indoor scene, generated by Midjourney V8.2 image-to-image" src="https://cdn-images-1.medium.com/max/1024/1*zOfzp84uVFgqdrHCGYlQqg.png" /><figcaption>Two-by-two grid of near-identical sage-green mugs still in the original indoor scene, generated by Midjourney V8.2 image-to-image</figcaption></figure><p>Settings: image-to-image with the mug result above as a data URL, aspect_ratio 3:2, 73 seconds, one output at 2048x1364.</p><p>The scene did not change: the mug stayed indoors in the original light. The task also returned one image containing a two-by-two grid, so each variation is only about 1024x682 and has to be cropped out. It stayed very close to the input, which may suit near-copies. This is one run with default settings, so test your own use case before relying on this endpoint.</p><h3>Prompting tips for Midjourney V8.2</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AjBKRTCt4-YluB6ve1MFGg.png" /><figcaption><em>Illustration generated with Midjourney V8.2 on Siray AI.</em></figcaption></figure><ol><li><strong>Pick </strong><strong>raw or high </strong><strong>stylize by goal.</strong> Our most literal results used raw, and our most expressive one used stylize 750. Midjourney&#39;s <a href="https://updates.midjourney.com/v8-alpha/">V8 alpha notes</a> also suggest --raw for photos or a plainer look.</li><li><strong>Put exact text in double quotes.</strong> Midjourney’s alpha notes tie better text to words “specified in quotes.” Keep it short: we tested two short lines, not paragraphs.</li><li><strong>Review all four candidates.</strong> Shape and composition varied noticeably, so the best image is often not the first.</li><li><strong>Plan for cropping.</strong> You pick from seven ratios, so resize afterward for fixed formats such as 1400x788.</li><li><strong>Add </strong><strong>negative_prompt for clean art.</strong> We passed &quot;text, letters, logo, watermark, people&quot; for the illustrations here and saw no stray text in the images we reviewed.</li></ol><h3>When to use Midjourney V8.2 (and when not to)</h3><p><strong>Good fit:</strong></p><ul><li>Concept art, mood boards and editorial illustration.</li><li>Painterly or strongly art-directed images.</li><li>Campaign exploration, where four candidates per task help.</li><li>Short display text, reviewed by a person.</li></ul><p><strong>Consider another model when:</strong></p><ul><li>You need precise edits to an existing image. See how <a href="https://blog.siray.ai/qwen-image-3-edit-shoe-colour/">Qwen Image 3 Edit recolors shoes</a>.</li><li>You need text on real packaging. See <a href="https://blog.siray.ai/gpt-image-2-5-flare-packaging-mockup/">GPT Image 2.5 Flare for packaging localization</a>.</li><li>You want a specific print style with composition checks. See <a href="https://blog.siray.ai/flux-kontext-pro-woodcut-posters/">FLUX Kontext Pro woodcut posters</a>.</li><li>Your product needs results within seconds. Our tasks took minutes.</li></ul><h3>FAQ</h3><h3>How much does the Midjourney V8.2 API cost per image?</h3><p>Siray AI lists text-to-image at about $0.021 per image. A task returned four images in our tests, and the listing does not say how that is counted, so budget up to $0.084 per task. Prices were checked on November 4, 2026.</p><h3>Is Midjourney V8.2 available via API?</h3><p>Yes. On Siray AI, use midjourney/midjourney-v8.2-t2i or midjourney/midjourney-v8.2-i2i through the async image endpoint with one API key.</p><h3>How long does Midjourney V8.2 take to generate images?</h3><p>Our seven text-to-image tasks took 192 to 372 seconds, including queue time. Our single image-to-image task took 73 seconds.</p><h3>Does Midjourney V8.2 support image-to-image?</h3><p>Yes, through midjourney/midjourney-v8.2-i2i, which takes one image as a URL or data URL. In our one test it stayed close to the input, ignored the requested scene change and returned a two-by-two grid.</p><h3>Can Midjourney V8.2 render text accurately?</h3><p>In our test, all four candidates spelled two short lines correctly when the text was in double quotes. Letter shapes and spacing were uneven in some, and we did not test long text.</p><h3>What sizes does Midjourney V8.2 return?</h3><p>You set one of seven aspect ratios rather than a pixel size. Our 3:2 tasks returned 1344x896 and our 16:9 tasks usually returned 1456x816, with one at 1024x573.</p><h3>Try Midjourney V8.2 on Siray AI</h3><p>Create a free account, get an API key and run the lighthouse prompt above. Open the <a href="https://console.siray.ai/model-detail/midjourney/midjourney-v8.2-t2i?ref=blog.siray.ai">Midjourney V8.2 model page on Siray AI</a> to start.</p><h3>Sources</h3><ul><li><a href="https://updates.midjourney.com/version-8-2/">Version 8.2</a>, Midjourney, published July 24, 2026, accessed November 4, 2026.</li><li><a href="https://updates.midjourney.com/edit-model-for-v8/">Edit Model for V8</a>, Midjourney, accessed November 4, 2026.</li><li><a href="https://updates.midjourney.com/v8-alpha/">V8 Alpha</a>, Midjourney, accessed November 4, 2026.</li><li><a href="https://docs.siray.ai/api-reference/model-api/midjourney-v8.2-t2i">Midjourney V8.2 t2i API reference</a>, Siray AI docs, accessed November 4, 2026.</li><li><a href="https://docs.siray.ai/api-reference/model-api/midjourney-v8.2-i2i">Midjourney V8.2 i2i API reference</a>, Siray AI docs, accessed November 4, 2026.</li><li><a href="https://docs.siray.ai/api-reference/task-status-image">Image task status API</a>, Siray AI docs, accessed November 4, 2026.</li><li>Siray AI model pricing API (GET /v1/models/pricing/{model}), queried November 4, 2026.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d0dcd1c4275d" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>