<?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 TokenLean on Medium]]></title>
        <description><![CDATA[Stories by TokenLean on Medium]]></description>
        <link>https://medium.com/@tokenlean?source=rss-332470537bdc------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*dmbNkD5D-u45r44go_cf0g.png</url>
            <title>Stories by TokenLean on Medium</title>
            <link>https://medium.com/@tokenlean?source=rss-332470537bdc------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 17:38:27 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@tokenlean/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[Gemini 4 Carbon doubles in price on Dec 1: what your agent bill does, and a 20-line fix]]></title>
            <link>https://medium.com/@tokenlean/gemini-4-carbon-doubles-in-price-on-dec-1-what-your-agent-bill-does-and-a-20-line-fix-c7258f034828?source=rss-332470537bdc------2</link>
            <guid isPermaLink="false">https://medium.com/p/c7258f034828</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[cost-optimization]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[TokenLean]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 14:05:38 GMT</pubDate>
            <atom:updated>2026-11-17T14:05:43.673Z</atom:updated>
            <content:encoded><![CDATA[<p>Gemini 4 Carbon doubles in price on Dec 1: what your agent bill does, and a 20-line fix</p><p><em>Disclosure: this post contains an affiliate link; I may earn a commission at no cost to you.</em></p><p><strong>TL;DR:</strong> Gemini 4 Carbon costs $2 in / $10 out per 1M tokens until Nov 30, then $4 / $20. Claude Fable 5.2 launched Nov 9 at $8 / $40 (20% below 5.1). Gemini 4 Hydrogen (Nov 12) is $1.50 / $7.50 flat. If you hard-coded one model, your December bill can double. Route by task instead.</p><h3>Prices (per 1M tokens, from the vendors’ announcements)</h3><h3>Worked example</h3><p>Assumption: an agent run uses 30K input and 5K output tokens, 2,000 runs a month. Carbon today: 0.03 x 2 + 0.005 x 10 = $0.11 per run, $220/month. On Dec 1: $0.22 per run, $440/month. Hydrogen: $0.0825 per run, $165/month. Fable 5.2: $0.44 per run, $880/month.</p><h3>Copy-paste: a tiny router</h3><pre>TIERS = { &quot;cheap&quot;: &quot;muse-ember&quot;, # classification, extraction &quot;default&quot;:&quot;gemini-4-hydrogen&quot;, # everyday coding/agent steps &quot;hard&quot;: &quot;gemini-4-carbon&quot;, # only when a cheaper tier failed } def pick(task): if task[&quot;kind&quot;] in (&quot;classify&quot;,&quot;extract&quot;,&quot;summarize&quot;): return TIERS[&quot;cheap&quot;] if task.get(&quot;retries&quot;,0) &gt;= 1 or task.get(&quot;risk&quot;)==&quot;high&quot;: return TIERS[&quot;hard&quot;] return TIERS[&quot;default&quot;]</pre><p>Send every call through one OpenAI-compatible gateway (LiteLLM works) so switching is a config edit, not a rewrite. Log tokens per tier weekly and check the share going to “hard”.</p><h3>Where to run it</h3><p>A small always-on box keeps the gateway and logs cheap and independent of your laptop. I run mine on a Hostinger VPS: <a href="https://go.nowto.link/hostinger?s=medium">Hostinger VPS</a>.</p><h3>Decision table</h3><h3>Sources</h3><p>Anthropic Fable/Mythos 5.2 announcement (Nov 9); Google Gemini 4 Carbon post (Nov 10); LiteLLM Gemini 4 Hydrogen post (Nov 12).</p><p><em>Full disclosure: some links on TokenLean are affiliate links. If you buy through them I may earn a commission at no extra cost to you. I only recommend tools I’d use; prices change, so verify on the vendor pages.</em></p><p><em>Originally published at </em><a href="https://tokenlean.wordpress.com/2026/11/16/gemini-carbon-price-doubles/"><em>http://tokenlean.wordpress.com</em></a><em> on November 16, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c7258f034828" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What a 1,000-step agent run costs on Meta’s new Muse Ember ($0.10 / $0.50 per 1M tokens)]]></title>
            <link>https://medium.com/@tokenlean/what-a-1-000-step-agent-run-costs-on-metas-new-muse-ember-0-10-0-50-per-1m-tokens-ae988163088a?source=rss-332470537bdc------2</link>
            <guid isPermaLink="false">https://medium.com/p/ae988163088a</guid>
            <category><![CDATA[agents]]></category>
            <category><![CDATA[hostinger]]></category>
            <category><![CDATA[cost]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[TokenLean]]></dc:creator>
            <pubDate>Tue, 10 Nov 2026 14:05:45 GMT</pubDate>
            <atom:updated>2026-11-12T14:06:34.559Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*i_q2SmYmkhCX-zCh" /></figure><p>What a 1,000-step agent run costs on Meta’s new Muse Ember ($0.10 / $0.50 per 1M tokens)</p><p><em>Disclosure: this post contains an affiliate link to Hostinger; I may earn a commission at no cost to you.</em></p><p><strong>TL;DR:</strong> Meta released Muse Ember on Nov 10, 2026, a small model aimed at the repetitive steps in agent loops (read a file, run a test, extract a field, call a tool). At $0.10 input, $0.004 cached input and $0.50 output per 1M tokens, a 1,000-step loop costs cents. Below: the math, a copy-paste cost calculator, and a routing rule for when to escalate to a bigger model.</p><h3>Prices (as published Nov 9–10, 2026)</h3><p>Sources: VentureBeat and Gadgets360 on Muse Ember; andrew.ooo price table for GPT-6.1 Terra. Anthropic also announced Claude Fable 5.2 / Mythos 5.2 with a 20% API cost cut, so check your frontier tier too.</p><h3>Worked example</h3><p>Assumptions (mine, change them): each step sends 6,000 input tokens, of which 5,000 are a cached prefix, and returns 300 output tokens.</p><pre>def run_cost(steps, in_tok, cached_tok, out_tok, p_in, p_cached, p_out): fresh = in_tok - cached_tok per_step = (fresh*p_in + cached_tok*p_cached + out_tok*p_out) / 1e6 return steps * per_step ember = run_cost(1000, 6000, 5000, 300, 0.10, 0.004, 0.50) terra = run_cost(1000, 6000, 0, 300, 1.00, 1.00, 5.00) # no cache discount assumed print(f&quot;Ember ${ember:.3f} Terra ${terra:.2f}&quot;)</pre><p>Result: about <strong>$0.27 on Ember vs $7.50 on Terra</strong> per 1,000 steps (Terra without caching, so this flatters Ember; with caching the gap narrows but stays large).</p><h3>Routing rule: cheap by default, escalate on failure</h3><pre>CHEAP, BIG = &quot;muse-ember-1.0&quot;, &quot;your-frontier-model&quot; def pick(task): if task.kind in {&quot;read&quot;,&quot;extract&quot;,&quot;run_tests&quot;,&quot;tool_call&quot;} and task.retries == 0: return CHEAP return BIG # planning, failed retries, ambiguous specs</pre><p>Log retries per task type for a week; if Ember’s retry rate on a step type exceeds ~20%, move that type to the bigger model.</p><h3>Where to run it</h3><p>A router like this is a tiny always-on process. A small VPS is enough; I use Hostinger for this kind of thing: <a href="https://go.nowto.link/hostinger?s=medium">Hostinger VPS</a>. Keep API keys in env vars, not in the repo.</p><h3>Decision table</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*2Fu-525fCVQs4nvy" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*XJj9BpEeMCuR-O96" /></figure><p>High-volume, simple, tool-shaped steps</p><p>Muse Ember</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*U9OyqJYifpvKyg94" /></figure><p>Planning, long reasoning, code design</p><p>Frontier model</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*S0pHZHeHokWqvd86" /></figure><p>Context above 200K tokens</p><p>Check the 3x price tier first</p><p><em>Full disclosure: some links in this article are affiliate links. If you buy through them, TokenLean may earn a commission at no extra cost to you. Prices change; verify on the vendor pages before relying on them.</em></p><p><em>Originally published at </em><a href="https://tokenlean.wordpress.com/2026/11/10/muse-ember-agent-cost/"><em>http://tokenlean.wordpress.com</em></a><em> on November 10, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ae988163088a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[One cheap-model router on a $5 VPS: use this week’s price drops without rewriting your app]]></title>
            <link>https://medium.com/@tokenlean/one-cheap-model-router-on-a-5-vps-use-this-weeks-price-drops-without-rewriting-your-app-b42e8be338fe?source=rss-332470537bdc------2</link>
            <guid isPermaLink="false">https://medium.com/p/b42e8be338fe</guid>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[cost]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[TokenLean]]></dc:creator>
            <pubDate>Sat, 07 Nov 2026 14:40:42 GMT</pubDate>
            <atom:updated>2026-11-08T14:01:59.320Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*0rCW8b3IaId4Mc5C.png" /></figure><p><strong>TL;DR:</strong> Prices moved a lot this week (Nov 2–6, 2026). Instead of hard-coding one model, put a tiny OpenAI-compatible router in front of your app, send easy requests to a cheap model and hard ones to a frontier model, and change the mapping in one file when prices move again.</p><h3>What moved this week</h3><p>Prices are list prices quoted by those trackers; check the provider page before you commit. They change daily.</p><h3>The router (copy-paste)</h3><p>Run it: uvicorn router:app --host 0.0.0.0 --port 8000, then point your app&#39;s OpenAI base URL at http://your-host:8000/v1. Streaming is not handled here; add it if you need it.</p><h3>Worked cost example (assumptions stated)</h3><p>100,000 requests a month, 1,500 input and 400 output tokens each = 150M input, 40M output tokens.</p><ul><li>All on Kimi K3.1 Pro: 150 x $2 + 40 x $10 = <strong>$700</strong></li><li>80% cheap (gpt-oss-120b) / 20% smart: cheap = 120 x $0.023 + 32 x $0.083 = $5.42; smart = 30 x $2 + 8 x $10 = $140; total <strong>about $145</strong></li></ul><p>Quality depends on how well your “hard” rule matches reality; log both routes for a week and compare.</p><h3>Where to host it</h3><p>The router is a tiny always-on process, so a small VPS is enough. I’d use <a href="https://go.nowto.link/hostinger?s=devto">Hostinger VPS</a> for this: put it behind Caddy for HTTPS, keep keys in environment variables, and restrict access with a shared token.</p><h3>Decision table</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VkHXzBqXJ0yvqHQV.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Nu7kf1QJquOLMEld.png" /></figure><p>Prices change weekly</p><p>Router with a ROUTES dict</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DRL6sIYiUatu4U5u.png" /></figure><p>Need data to stay in one region</p><p>Check provider residency terms first</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*SpCi9ZmIjZcHC5ea.png" /></figure><p>Sensitive prompts</p><p>Self-host open weights instead</p><h3>Sources</h3><ul><li>overfeed.news OpenRouter pricing snapshot (2026–11–05)</li><li>thismonthinai.com November 2026 (Together AI Kimi K3.1 Pro pricing)</li><li>promptunit.ai AI API pricing (Nov 2026)</li></ul><p><em>Full disclosure: TokenLean uses affiliate links, including the Hostinger link above. If you buy through them I may earn a commission at no extra cost to you. I only recommend tools I’d consider for the job described. Prices are third-party reported and may be out of date.</em></p><p><em>Originally published at </em><a href="https://dev.to/tokenlean/one-cheap-model-router-on-a-5-vps-use-this-weeks-price-drops-without-rewriting-your-app-3eoo"><em>https://dev.to</em></a><em> on November 7, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b42e8be338fe" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Rent a GPU for open-weight models: do the VRAM math first (70B on one card, about $7 to test)]]></title>
            <link>https://medium.com/@tokenlean/rent-a-gpu-for-open-weight-models-do-the-vram-math-first-70b-on-one-card-about-7-to-test-84c61c88e1a3?source=rss-332470537bdc------2</link>
            <guid isPermaLink="false">https://medium.com/p/84c61c88e1a3</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[gpu]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[TokenLean]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 14:09:56 GMT</pubDate>
            <atom:updated>2026-11-05T14:32:02.264Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*OSts6tCHW5perV-q" /></figure><p>Rent a GPU for open-weight models: do the VRAM math first (70B on one card, about $7 to test)</p><p><em>Disclosure: this post contains affiliate links; I may earn a commission at no cost to you.</em></p><h3>TL;DR</h3><p>Before you rent a GPU for an open-weight model, do the VRAM math. Weights take roughly parameters x bytes per weight, then add 15 to 25 percent for KV cache and runtime. A 70B model at 4-bit fits one 80 GB card. A 8B model fits a cheap 24 GB card. Rent by the second, stop the pod when idle, and compare against API pricing before committing.</p><h3>The VRAM calculator (copy-paste)</h3><pre>def vram_gb(params_b, bits=4, overhead=0.2): weights = params_b * bits / 8 # GB for weights return round(weights * (1 + overhead), 1) for name, p in [(&quot;8B&quot;, 8), (&quot;32B&quot;, 32), (&quot;70B&quot;, 70)]: print(name, {b: vram_gb(p, b) for b in (16, 8, 4)})</pre><p>Long contexts and many concurrent users grow the KV cache, so treat the table as a floor, not a promise.</p><h3>Launch an OpenAI-compatible endpoint with vLLM</h3><p>On a pod with a vLLM template (or any CUDA image):</p><pre>pip install vllm vllm serve &lt;your-huggingface-model-id&gt; \ --max-model-len 16384 \ --gpu-memory-utilization 0.90 \ --api-key $MY_KEY</pre><p>Then point any OpenAI SDK at http://&lt;pod-host&gt;:8000/v1. For quantized weights add --quantization awq (or load a GGUF with llama.cpp if you have a smaller card).</p><h3>What it costs (assumptions stated)</h3><p>Prices below are RunPod’s published Secure Cloud on-demand rates as seen on its pricing page in late October 2026, billed per second. They move often, so check the page.</p><p>Worked example: you test a 4-bit 70B model for 4 hours on an A100 80GB, then stop the pod. That is about $7. A pod left running all month at the same rate is about $1,290 (1.79 x 720 h). The habit that saves money is simply stopping the pod. If your traffic is bursty, serverless endpoints bill only while workers run, but their hourly rate is higher, so check your utilization first.</p><h3>Self-host or use an API?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*rRw_kFf6ttAjhaW2" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*jQ7jo1cFcGVQxthS" /></figure><p>Under a few hours of use per day, spiky</p><p>API, or serverless GPU</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*jy2J22vQWretmbNT" /></figure><p>Steady high volume, same model all day</p><p>Dedicated pod</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*jepCv88sG2oCqXEG" /></figure><p>Data must stay in your control</p><p>Rented GPU you configure</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*WkKmyTfyxTJTGLu-" /></figure><p>Fine-tuned or niche open-weight model</p><p>Rented GPU</p><h3>Try it</h3><p>If you want to test this without buying hardware, you can spin up a pod on <a href="https://go.nowto.link/runpod?s=medium">RunPod here </a>and be serving a model in about ten minutes. Start with the 24 GB tier for 8B models and only step up when the math above says you must.</p><h3>Checklist before you click deploy</h3><ul><li>Computed VRAM with the function above, plus context length</li><li>Picked 4-bit if quality is acceptable for your task</li><li>Set a calendar reminder to stop the pod</li><li>Attached a network volume so weights are not re-downloaded</li></ul><p>Source: RunPod pricing page (runpod.io/pricing), vLLM docs.</p><p><em>Full disclosure: TokenLean participates in affiliate programs, including RunPod’s. If you sign up or buy through links on this page, I may earn a commission at no extra cost to you. I only recommend tools I think are useful, and prices are quoted from public pages at the time of writing and may change.</em></p><p><em>Originally published at </em><a href="https://tokenlean.wordpress.com/2026/11/05/runpod-vram-math-rent-gpu-open-weight/"><em>http://tokenlean.wordpress.com</em></a><em> on November 5, 2026.</em></p><ul><li>Set a calendar reminder to stop the pod</li><li>Attached a network volume so weights are not re-downloaded</li></ul><p>Source: RunPod pricing page (runpod.io/pricing), vLLM docs.</p><p><em>Full disclosure: TokenLean participates in affiliate programs, including RunPod’s. If you sign up or buy through links on this page, I may earn a commission at no extra cost to you. I only recommend tools I think are useful, and prices are quoted from public pages at the time of writing and may change.</em></p><p><em>Originally published at </em><a href="https://tokenlean.wordpress.com/2026/11/05/runpod-vram-math-rent-gpu-open-weight/"><em>http://tokenlean.wordpress.com</em></a><em> on November 5, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=84c61c88e1a3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Murf vs ElevenLabs API: what 1 million characters of speech really costs (Nov 2026)]]></title>
            <link>https://medium.com/@tokenlean/murf-vs-elevenlabs-api-what-1-million-characters-of-speech-really-costs-nov-2026-75d6921544b4?source=rss-332470537bdc------2</link>
            <guid isPermaLink="false">https://medium.com/p/75d6921544b4</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[api]]></category>
            <category><![CDATA[text-to-speech]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[pricing]]></category>
            <dc:creator><![CDATA[TokenLean]]></dc:creator>
            <pubDate>Wed, 04 Nov 2026 14:11:17 GMT</pubDate>
            <atom:updated>2026-11-04T14:28:04.421Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*Pvs0XXgSgL8BHeCk" /></figure><p>Murf vs ElevenLabs API: what 1 million characters of speech really costs (Nov 2026)</p><p><em>Disclosure: this post contains affiliate links; I may earn a commission at no cost to you.</em></p><p><strong>TL;DR:</strong> On list prices, Murf’s Falcon 2.1 is $0.01 per 1,000 characters and ElevenLabs’ cheapest API tier is $0.04. At 10 million characters a month that is $100 against $400. The gap is real, but price is only one of four things to check before you switch.</p><h3>The rate card (list prices, pay-as-you-go API)</h3><p>Sources: murf.ai/pricing and elevenlabs.io/pricing/api, as checked 4 Nov 2026. Murf’s purchased credit does not expire and the free trial gives $10 of API credit with no card. Vendors change prices often, so re-check before committing.</p><h3>Worked cost example</h3><p>Assumption: about 1,000 characters is roughly one minute of speech. Say your app speaks 10,000 minutes a month, so 10M characters.</p><h3>A fair A/B test in 20 lines</h3><p>Do not trust anyone’s demo page, including mine. Render the same ten of your own lines with both and listen blind.</p><pre>import requests, pathlib LINES = pathlib.Path(&quot;lines.txt&quot;).read_text().splitlines()[:10] def murf(text, i): r = requests.post(&quot;https://api.murf.ai/v1/speech/generate&quot;, headers={&quot;api-key&quot;: MURF_KEY}, json={&quot;text&quot;: text, &quot;voiceId&quot;: &quot;YOUR_VOICE_ID&quot;, &quot;modelVersion&quot;: &quot;FALCON&quot;}) # check Murf&#39;s docs for the exact response field holding the audio URL open(f&quot;murf_{i}.mp3&quot;,&quot;wb&quot;).write(requests.get(r.json()[&quot;audioFile&quot;]).content) def eleven(text, i): r = requests.post(&quot;https://api.elevenlabs.io/v1/text-to-speech/YOUR_VOICE_ID&quot;, headers={&quot;xi-api-key&quot;: ELEVEN_KEY}, json={&quot;text&quot;: text, &quot;model_id&quot;: &quot;eleven_flash_v2_5&quot;}) open(f&quot;eleven_{i}.mp3&quot;,&quot;wb&quot;).write(r.content) for i, t in enumerate(LINES): murf(t, i); eleven(t, i)</pre><p>Field names and model IDs can change; confirm both against each vendor’s current API docs before running.</p><h3>Decision table</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*0gNyvgYNSTOQaeaw" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*fUWPQm6TfPEIpZO3" /></figure><p>Lowest per-character cost for agents or bulk narration</p><p>Murf Falcon</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*g-6a8cxTDZJOTr2p" /></figure><p>Studio-quality voiceover at a moderate price</p><p>Murf Gen2</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*LDa1DwyU2g_mbvre" /></figure><p>Widest language and expressiveness range, voice cloning on cheap plans</p><p>ElevenLabs</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*rY1jIxfkpQXz_8sZ" /></figure><p>Not sure</p><p>Run the A/B above, it costs under a dollar</p><p>Four things to check beyond price: voice quality for your language, latency in your region, request-size limits (Murf caps requests at around 3,000 characters per third-party reports), and licensing for commercial use.</p><p><strong>Try Murf:</strong> <a href="https://go.nowto.link/murfhark?s=medium">Murf API free trial</a><br><strong>Or stay with ElevenLabs:</strong> <a href="https://go.nowto.link/elevenhark?s=wp">ElevenLabs</a></p><h3>Sources</h3><p><em>Full disclosure: TokenLean uses affiliate links for Murf and ElevenLabs. If you sign up through them I may earn a commission at no extra cost to you. I only recommend tools I think fit the job described.</em></p><p><em>Originally published at </em><a href="https://tokenlean.wordpress.com/2026/11/04/murf-vs-elevenlabs-api-cost/"><em>http://tokenlean.wordpress.com</em></a><em> on November 4, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=75d6921544b4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Switching an OpenAI Realtime voice agent to ElevenLabs: the 2-line change, the real cost per call…]]></title>
            <link>https://medium.com/@tokenlean/switching-an-openai-realtime-voice-agent-to-elevenlabs-the-2-line-change-the-real-cost-per-call-2de7e6ff0a4e?source=rss-332470537bdc------2</link>
            <guid isPermaLink="false">https://medium.com/p/2de7e6ff0a4e</guid>
            <category><![CDATA[voice-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[javascript]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[eleven-labs]]></category>
            <dc:creator><![CDATA[TokenLean]]></dc:creator>
            <pubDate>Tue, 03 Nov 2026 16:22:54 GMT</pubDate>
            <atom:updated>2026-11-03T16:49:08.421Z</atom:updated>
            <content:encoded><![CDATA[<h3>Switching an OpenAI Realtime voice agent to ElevenLabs: the 2-line change, the real cost per call, and the WebRTC catch</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*cUIs4TaEzWrRvD97" /></figure><p>Switching an OpenAI Realtime voice agent to ElevenLabs: the 2-line change, the real cost per call, and the WebRTC catch</p><p><em>Disclosure: this post has affiliate links. If you sign up through them I may earn a commission, at no extra cost to you.</em></p><p><strong>TL;DR:</strong> ElevenLabs shipped Eleven Realtime on Oct 20, 2026: two speech-to-speech models (v1 Mini and v1 Max) behind an API that speaks OpenAI’s Realtime protocol. If your voice agent already uses @openai/agents-realtime over WebSocket, the switch is a new endpoint URL and a new model string. If you&#39;re on WebRTC, it&#39;s a transport rewrite first. Cost is $0.08 per connected minute plus LLM tokens, and the per-minute part dominates.</p><h3>What actually shipped</h3><ul><li><strong>Eleven Realtime v1 Mini</strong> (eleven_realtime_v1_mini): the default. Built for fast, natural conversation: assistants, tutors, characters.</li><li><strong>Eleven Realtime v1 Max</strong> (eleven_realtime_v1_max): reasons before it answers. Built for support, lead qualification and multi-step tool calling.</li><li>Both cover 90+ languages and any of the 17,000+ voices in the ElevenLabs library.</li><li>Under the hood it’s a <em>cascaded</em> stack (speech recognition, turn-taking, an LLM, and a voice), tuned end to end and exposed as one model ID. The practical upside: every step produces text, so you can see what the agent heard versus what it decided to say.</li></ul><pre>Browser mic ──PCM16 24kHz──► wss://api.elevenlabs.io/v1/convai/realtime │ ┌──────────┴───────────┐ │ ASR → turn-taking → │ one model ID: │ LLM → TTS (tuned as │ eleven_realtime_v1_mini │ a single system) │ eleven_realtime_v1_max └──────────┬───────────┘ Speaker ◄──PCM16 audio───────────┘ (+ text at every step for logs/guardrails)</pre><h3>The numbers ElevenLabs published</h3><p>Benchmarks are from Artificial Analysis as reported in ElevenLabs’ launch post; prices from the ElevenLabs API pricing page (checked Nov 3, 2026).</p><p>The one line that should drive your model choice: <strong>tool use, 60% vs 87%.</strong> If your agent calls functions (book a slot, look up an order), Mini will drop roughly 4 in 10 tasks on that benchmark. If it mostly chats, Mini is the better-sounding, cheaper pick.</p><h3>The migration: what changes in your code</h3><p>ElevenLabs’ docs say the API uses the same client and server events over the same WebSocket transport as OpenAI’s Realtime API. In practice there are two changes in the browser, plus one on your server.</p><p>You’ll need an ElevenLabs API key for step 1 ( <a href="https://go.nowto.link/elevenhark?s=medium">create one here</a>; the free tier is enough to test).</p><p><strong>1. Server: mint a short-lived client secret (never ship your API key to the browser).</strong></p><pre>import os from elevenlabs.client import ElevenLabs from elevenlabs import (CreateClientSecretRequest, ExpiresAfter, OpenAiSessionConfigInput, OpenAiAudioConfigInput, OpenAiAudioInputConfigInput, OpenAiAudioInputFormat, OpenAiAudioOutputConfigInput, OpenAiAudioOutputFormat, OpenAiTurnDetection) el = ElevenLabs(api_key=os.environ[&quot;ELEVENLABS_API_KEY&quot;]) secret = el.conversational_ai.realtime.create_client_secret( request=CreateClientSecretRequest( session=OpenAiSessionConfigInput( type=&quot;realtime&quot;, model=&quot;eleven_realtime_v1_mini&quot;, # or eleven_realtime_v1_max instructions=&quot;You are a concise, friendly voice assistant.&quot;, output_modalities=[&quot;audio&quot;], audio=OpenAiAudioConfigInput( input=OpenAiAudioInputConfigInput( format=OpenAiAudioInputFormat(type=&quot;audio/pcm&quot;, rate=24000), turn_detection=OpenAiTurnDetection(type=&quot;semantic_vad&quot;, eagerness=&quot;medium&quot;), ), output=OpenAiAudioOutputConfigInput( format=OpenAiAudioOutputFormat(type=&quot;audio/pcm&quot;), voice=&quot;brian&quot;), ), ), expires_after=ExpiresAfter(anchor=&quot;created_at&quot;, seconds=600), ) ) # return only secret.value to the browser</pre><p><strong>2. Browser: same OpenAI SDK, new URL and model.</strong></p><pre>import { OpenAIRealtimeWebSocket, RealtimeAgent, RealtimeSession } from &quot;@openai/agents-realtime&quot;; const { value: clientSecret } = await fetch(&quot;/api/token&quot;).then(r =&gt; r.json()); const agent = new RealtimeAgent({ name: &quot;support&quot;, instructions: &quot;Keep replies short.&quot; }); const transport = new OpenAIRealtimeWebSocket({ url: &quot;wss://api.elevenlabs.io/v1/convai/realtime&quot;, // ← change #1 }); const session = new RealtimeSession(agent, { transport, model: &quot;eleven_realtime_v1_mini&quot;, // ← change #2 config: { outputModalities: [&quot;audio&quot;], audio: { output: { voice: &quot;brian&quot; } } }, }); await session.connect({ apiKey: clientSecret });</pre><h3>The catches nobody puts in the headline</h3><ol><li><strong>WebSocket only.</strong> If your app uses the SDK’s WebRTC transport, you have to move to WebSocket first.</li><li><strong>You handle the audio.</strong> The OpenAI SDK only wires mic and speaker automatically over WebRTC. Over WebSocket you capture and play PCM16 at 24 kHz yourself (wavtools works; it&#39;s what OpenAI&#39;s own WebSocket examples use).</li><li><strong>Browser flow only, for now.</strong> The docs say server-side OpenAI Realtime SDK docs aren’t published yet, so phone-bridge setups that run the SDK on a server are untested territory.</li><li><strong>The API is separate from ElevenAgents.</strong> Phone numbers, WhatsApp, knowledge base, guardrails and testing live in ElevenAgents. If you need those, build there and pick a Realtime model from the dropdown instead.</li><li><strong>Burst pricing.</strong> Connected audio above your plan’s concurrency limit bills at $0.16/min, double the normal rate. Concurrency runs from 4 calls (free/pay-as-you-go) to 40 (Business).</li></ol><h3>What a 5-minute call costs (worked estimate)</h3><p>The connected-audio fee is fixed: 5 × $0.08 = <strong>$0.40</strong>. Tokens depend on your prompt and turn count. Here’s an illustrative support call with my assumptions spelled out: 30 turns, about 3,000 tokens of context re-sent per turn (90k input tokens total), and about 60 output tokens per reply (1.8k output).</p><p>Two takeaways:</p><ul><li><strong>On Mini, tokens are noise; minutes are the bill.</strong> Hang up faster (end-call tool, tight prompts, no long hold music) and you save more than any prompt golf would.</li><li><strong>On Max, input tokens are the lever.</strong> At $1.80/M input, a bloated 3k-token system prompt re-sent every turn is about a quarter of the call cost. Cut it to 800 tokens and the Max call drops toward $0.45.</li></ul><h3>Copy-paste: a voice-agent system prompt that keeps calls short</h3><pre>You are {brand}&#39;s phone assistant. Goal: resolve the caller&#39;s issue in under 3 minutes. Rules: - Answer in 1-2 sentences. Never read lists longer than 3 items aloud. - Ask one question at a time. - If you need data, call the tool immediately; say &quot;one sec&quot; while it runs. - When the issue is resolved, confirm in one sentence and call end_call. - If the caller asks for a human or you fail twice, transfer.</pre><h3>Which one should you pick?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*GDkG4UZACV1V-UJH" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*m4XQDQ97_fSibAoo" /></figure><p>Chats, tutors, plays a character</p><p><strong>v1 Mini</strong>: cheapest, best turn-taking score</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*hzKzAqZQvMJC9SUH" /></figure><p>Calls tools / follows a script / qualifies leads</p><p><strong>v1 Max</strong>: 87% vs 60% on tool tasks</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*dDI4cH9uWMAyfFWQ" /></figure><p>Needs phone, WhatsApp, a knowledge base</p><p><strong>ElevenAgents</strong> with a Realtime model selected</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*GBoh4VijozIC-kN0" /></figure><p>Is an existing OpenAI Realtime app on WebSocket</p><p><strong>Realtime API</strong>: swap URL + model, test today</p><p>If you want to try it, the free tier includes roughly 14 minutes of Realtime v1 Mini, enough to run your existing client against it this afternoon. <a href="https://go.nowto.link/elevenhark?s=medium"><strong>Get an ElevenLabs API key →</strong></a></p><p>Sources: ElevenLabs launch post (elevenlabs.io/blog/realtime), OpenAI Realtime compatibility guide (elevenlabs.io/docs/eleven-agents/guides/openai-realtime-compatibility), ElevenLabs API pricing (elevenlabs.io/pricing/api).</p><p><em>Affiliate disclosure: TokenLean participates in affiliate programs, including ElevenLabs’. Some links in this article are affiliate links, which means I may earn a commission if you click and make a purchase, at no additional cost to you. I only recommend tools I’d use in a real build, and the facts above come from the vendor’s public docs and pricing pages. Prices and benchmarks change; check the official pages before you commit.</em></p><p><em>Originally published at </em><a href="https://tokenlean.wordpress.com/2026/11/03/elevenlabs-realtime-openai-migration-cost/"><em>http://tokenlean.wordpress.com</em></a><em> on November 3, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2de7e6ff0a4e" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>