Sketch, templates, image comments and prompt sharing all do the same thing: they take work away from the words. Here is what that means if you make images for a living.

The launch everyone read as a quality update

When OpenAI shipped ChatGPT Images 2.5 on September 8, most of the coverage led with the obvious things. Sharper detail, more natural lighting, better at keeping a face recognisable across edits, and generation latency cut by up to half compared with Images 2.0. Two new API models, Flare for everyday volume and Sunburst for precision work. All true, and all the kind of thing every image model launch says.

I read the announcement twice before I noticed what I think is the more interesting story. Four of the new features have nothing to do with image quality. They are Sketch, templates, comments placed directly on an image, and the option to share the prompt behind an image so other people can reuse it.

Look at those four together and they point the same way. Every one of them takes work away from the text prompt.

For three years the prompt has been treated as the craft. People sold prompt packs, guarded their “secret” phrasing, and argued about whether “cinematic lighting” still did anything. Images 2.5 is the clearest signal yet that the company behind the most-used image tool does not think the prompt should carry that much weight.

Sketch replaces description with drawing

Sketch lets you type “@Sketch” in ChatGPT and draw directly in the conversation. OpenAI’s examples are a room layout, the outline of an outfit, a doodle. You add a line about style, and the model turns the rough drawing into a finished image.

This matters because spatial layout is the thing text has always been worst at. Researchers have documented for years that image models struggle with simple positional instructions. Ask for a dog to the left of a cat and you get a coin toss. I have lost more evenings than I would like to admit rewriting “in the foreground, slightly off-centre, facing away from the window” in six different ways, when what I actually wanted could have been drawn in four seconds with a rectangle and an arrow.

A sketch resolves that ambiguity instantly. You are no longer describing a composition. You are showing it.

There is some useful research on how the two modes differ. A CHI 2024 study comparing sketch-guided and prompt-guided generative tools in a design task found they did different jobs. Prompts were most useful for sparking initial ideas. Sketches were where designers pinned those ideas down and checked whether they worked. That matches my own experience almost exactly. Words are good for “what if”, drawing is good for “like this”.

So Sketch does not make the prompt obsolete. It takes the part of the prompt that was always doing the worst job, spatial arrangement, and hands it to a better tool.

Templates replace structure with a starting point

The second feature is templates for common formats. OpenAI names flyers, product photos, posters and merch. You pick one and fill in what the image needs to say and how it should look.

If you have ever written a long prompt for a poster, you know how much of it is scaffolding. Aspect ratio, where the headline goes, how much empty space to leave, what counts as the hero element. None of that is creative. It is the boring structural layer that has to be right before the interesting choices mean anything.

A template absorbs that layer. What is left for you to write is the part that was always yours: the message, the mood, the specific details that make it your flyer and not a generic one.

I think this is quietly the most important change for beginners. The biggest gap between a first-week user and someone experienced was never taste. It was knowing which boring structural words to include. Templates close most of that gap on day one.

Comments replace re-explaining with pointing

The third feature is the one I expect to use most. You can now place comments directly on a specific part of an image and ask for a change there.

Before this, editing meant writing a sentence that located the thing you wanted changed. “The mug on the left side of the table, not the one near the laptop.” Then hoping the model found the same mug you meant. Half the frustration of iterative editing was never about the edit. It was about the model misunderstanding which part of the image you were talking about.

Pointing is how people give feedback to each other. An art director does not write a paragraph about where the logo is. They circle it. Images 2.5 finally lets you circle it.

This pairs with the other editing improvement in the release, better consistency across multiple rounds of changes. If earlier edits stay put while you make new ones, and you can point at exactly what you want changed, the editing loop starts to look much more like working with a person.

Prompt sharing makes the prompt a commodity

The fourth feature is the one that says the most about where this is heading. When you share an image, you can now include the prompt that made it, and someone else can run that same idea with their own photos and details. OpenAI’s own example is an “80s flashback” portrait prompt that it describes as going viral.

Think about what that does to the idea of a secret prompt. If the platform itself makes prompts one tap to copy, and the most popular ones spread as trends, then a prompt stops being an asset you own. It becomes something closer to a meme format. Useful, fun, and worth almost nothing on its own, because everyone has it.

I do not think this is bad news. It is honest news. The prompt-pack economy was always selling something that was going to be free eventually, and now it is.

What you cannot copy with one tap is everything around the prompt. Which reference photos you chose. Which of forty outputs you kept. The three edits you made afterwards because the first version felt slightly off. The reason you were making the image at all.

What the prompt is still for

I want to be careful not to overstate this. The prompt is not dead, and Images 2.5 still reads text. There are a few things only words can do well.

Words carry intent that cannot be drawn. “Make it feel like the last warm evening before autumn” is not a shape. It is a mood, and a sketch cannot hold it.

Words carry constraints. Brand colours, what must not appear, the exact copy on a flyer, the audience it is for. Templates give you a place to put those, but you still have to know what they are.

And words are still where ideas start. That CHI finding is worth repeating: prompts were most useful at the very beginning, when the goal was to generate options rather than finish one. If you open Sketch before you know what you want, you will probably draw the first thing that comes to mind and then refine it into something competent and forgettable.

So the new shape of the prompt is shorter and more about meaning. Less “85mm lens, shallow depth of field, rule of thirds, soft rim light”. More “a quiet morning, someone who has just made a decision”. The technical vocabulary moves into templates, drawings and comments. The part that stays in the words is the part that was always the actual creative choice.

How I am changing my own process

Here is what I have started doing differently since trying the new features.

I write the first prompt as a single sentence about intent, with no technical terms at all. If I cannot say what the image is for in one sentence, I am not ready to make it.

I sketch the layout the moment the composition matters, instead of trying to describe it. Boxes and arrows are enough. The drawing does not need to be good, it needs to be clear.

I use a template whenever the format is standard. Posters and product shots are solved problems. There is no craft in re-solving them.

I edit by pointing, not by describing. If I catch myself writing “the one on the left”, I stop and place a comment instead.

And I no longer save prompts as if they were valuable. I save the reference images, the rejected outputs, and a note on why I picked the one I kept. That is the part of the work that is actually mine.

The bigger shift

Every creative tool eventually hides its mechanics. Early photographers had to understand chemistry. Early desktop publishing meant typing codes to set a font. The mechanics never disappear, but they stop being the thing that separates good work from bad.

The prompt was the mechanics of AI image-making, and for a while it looked like the craft itself. Images 2.5 is the moment the tool started hiding it. Drawing, templates, pointing and copyable prompts all push the same way: less effort spent explaining to the machine, more effort spent deciding what is worth making.

That is a better deal for anyone who was in this for the images rather than the incantations.

Your Turn

Have you tried Sketch or image comments yet? I am curious whether drawing the layout has actually changed your results, or whether you still end up back in the text box. And if you used to keep a folder of prompts you were proud of, what are you keeping now instead? Tell me in the responses.

Source: OpenAI, “Introducing ChatGPT Images 2.5” (September 8, 2026). Research reference: “The Impact of Sketch-guided vs. Prompt-guided 3D Generative AIs on the Design Exploration Process,” CHI 2024.

Cheaper, faster, and better on almost every chart. I read the footnotes, the independent numbers and the migration notes to find where the headline stops being true.

Anthropic released Claude Opus 5.5 on September 22, and the launch post makes three big claims. It performs at the level of Fable 5.1, Anthropic’s more expensive model, “on most work.” It costs 40% less to run than Opus 5. And it generates output more than 30% faster.

Launch claims are written to be quoted, so I spent the first few days on the less exciting part: the footnotes, the pricing page, the developer migration notes, and the first independent evaluation. Most of the headline holds up. A few parts of it only hold under specific conditions, and one domain moved backwards.

Here’s my read, with the verdict first.

The short answer

If you use Claude through a subscription, switch. You lose nothing, you get more usage out of the same plan, and the writing is noticeably easier to work with.

If you run Opus 5 through the API, switch too, but don’t just change the model name. The 40% saving depends on a setting most integrations will have pinned, and four things that work on Opus 5 now return errors.

If your work is legal research, medical coding or tax, test before you move. That’s the main area where an independent evaluator measured Opus 5.5 doing worse than the model it replaces.

What got cheaper, and the condition attached

The per-token prices dropped by a fifth across the board. Input is $4 per million tokens instead of $5, output is $20 instead of $25, and cache writes are $5 instead of $6.25.

The bigger change is cache reads, which fall from $0.50 to $0.20 per million, a 60% cut. That matters more than it sounds, because in agentic and coding work the model re-reads the same large context over and over. Anthropic says cache reads make up the majority of those costs. So the biggest discount landed on the biggest line of an agent’s bill.

The 40% figure has a condition attached, though. Anthropic’s wording is that “at default settings” Opus 5.5 costs 40% less “on typical workloads.” The default effort level on Opus 5.5 is medium. On Opus 5 it was high. So the headline compares Opus 5.5 thinking less against Opus 5 thinking more, and Anthropic’s own guidance says medium on the new model matches or beats high on the old one.

That’s fair for most people. It’s a trap for anyone who tuned an integration on Opus 5. According to Anthropic’s migration notes, at the same named effort level the new model tends to think more per turn than Opus 5 did. Keep your old “high” setting, swap the model name, and you’ll get the per-token discount but lose some or all of the token savings. One analysis put the pure price effect at 20% with no caching and about 29% for a cache-heavy agent task. The rest of the 40% has to come from fewer tokens per task.

What got better

On Anthropic’s own table, Opus 5.5 beats Opus 5 on every benchmark listed. Some of the jumps are large:

On Terminal-Bench 4.0, a test of multi-step work in a command line, it goes from 52.3% to 66.4%. On GDPval-AA, which scores real professional tasks across 44 occupations, it goes from 1708 to 1846 Elo. On OSWorld, a computer-use test, it moves from 74.0% to 81.8%. And on an agentic scientific research benchmark it roughly doubles, from 29.0% to 58.7%.

The examples are more telling than the scores. Asked to cut load times across every page of a web app, Opus 5.5 succeeded 39 times out of 40, while Opus 5 made smaller improvements that also changed how the app behaved. In a research test where every figure and quote was checked against sources, 16 of 18 Opus 5.5 reports passed, and neither Opus 5 nor Fable 5.1 passed in any attempt. That second result matters more to me than any leaderboard, because an invented number is exactly the failure that makes people stop trusting these tools.

Two caveats come straight from Anthropic, and I think they deserve more attention than they’re getting. First, the scores in the headline table were run at max effort, not at the medium default where the cost saving is measured. The two best numbers in the launch come from two different settings. Second, Anthropic itself writes that at this level of capability “benchmark margins have become a less reliable guide to real-world differences.” When the company selling the model tells you to discount its own chart, take it at its word.

What the independent numbers say

Vals AI, an evaluation firm, ran Opus 5.5 through its own benchmark suite on launch day. As I write this, its overall index ranks Opus 5.5 first of 65 models at 69.69%, ahead of Fable 5.1 at 68.83% and Opus 5 at 67.21%. It took first place on six individual leaderboards, including Terminal-Bench 4.0 and a perfect score on a proof benchmark.

But the same page shows where it slipped. On Harvey’s Legal Agent Benchmark it scored 3.75%, ranking 31st of 64. On MedCode, a medical coding test, it landed below Opus 5. An early launch-day write-up reported Vals figures for Opus 5 against Opus 5.5 on tax and public benefits work that also went down, not up. So the pattern is big gains on long, iterative technical work and small losses on some retrieval-heavy professional tasks.

Two more details worth knowing. Vals ran Opus 5.5 with older Claude models as fallbacks for refused requests, and on one site-reliability benchmark 83% of tasks were handled with fallback help. Counting those as failures drops that score from 33.59% to 5.34%. And that same early write-up had Opus 5.5 fourth on the Vals index, just below Opus 5. The page now shows it first. Launch-week numbers are still moving, and any article quoting them, this one included, has a shelf life.

The change that matters most to me: the writing

This is the part I care about most, because most of what I do with these models is writing and editing rather than code.

Opus 5 had a recognisable voice, and not in a good way. Dense prose, idiosyncratic phrasing, the most useful sentence buried a paragraph down. Anthropic says this was one of the most common complaints about it, and the launch post addresses it directly: Opus 5.5 “puts the most important information up front, is less likely to use jargon or idiosyncratic phrases, and follows the writing rules you give it.”

The side-by-side example in the announcement shows the difference well. Asked to explain a billing bug, Opus 5 opens with the commit and the code. Opus 5.5 opens with the answer: how much of the revenue drop comes from a free-tier change, how much from the bug, and what the bug does. Same facts, different order, and the second one is what you’d want from a colleague.

Partner reports point the same way. Box says Opus 5.5 used a third of the tokens Opus 5 needed, with answers 40% less verbose and no loss of accuracy. For anyone who trims a model’s output before pasting it anywhere, that’s the whole point.

The part I’d test first is “follows the writing rules you give it.” If you keep a style guide in your prompts, and you should, this is a claim you can check in an afternoon.

What breaks if you’re a developer

If you only use the chat app, skip this section. If you’ve built on the API, four things that work on Opus 5 now return errors:

Thinking can’t be switched off. Requests that disable it, or set a manual thinking budget, are rejected, and effort becomes the only control. Forced tool use is gone, so an integration that makes the model call a specific tool has to be reworked. Earlier reasoning is now tied to the exact conversation it came from, so apps that edit the system prompt or earlier messages mid-conversation will fail on replay. And the older computer-use tool version is rejected on the Claude API and Google Cloud.

There’s also a change that fails silently. The short progress notes the model writes between tool calls now come back as thinking blocks, which are empty by default. An app that shows those notes to users as status updates will just go quiet, with no error.

And one behaviour change for anyone running unattended agents: on long tasks, Opus 5.5 sometimes ends a turn with a progress report instead of a tool call. A loop that treats that as “done” will stop halfway.

Finally, Opus 5.5 ships with Fable-class safeguards for cybersecurity and biology. Fixing bugs in your own code is unaffected, but most cybersecurity tasks are quietly routed to an older model unless you’re in Anthropic’s verification program.

So who should switch?

Subscribers: yes, today. Anthropic raised the five-hour usage limits on Pro, Max and Team plans at the same time, and because Opus 5.5 is cheaper to run, it goes further inside them. You also get a one-off limit reset you can save for a long session. Don’t spend it on day one.

API users on Opus 5: yes, but properly. Change the model name in a branch, fix the four breaking changes, delete whatever effort setting you inherited from Opus 5, and start at medium. Then compare two or three effort levels on your own traffic and measure cost per completed task, not cost per token.

Legal, medical coding and tax workloads: run your own evaluation first. The independent data shows regressions here, and a newer model is not automatically a better one for your specific job.

Fable 5.1 users: this may be the biggest win of all. Opus 5.5 lists at 40% of Fable’s per-token price and beats it on Anthropic’s published table, although Anthropic says the real gap is smaller than the scores suggest. Test it on your hardest tasks, because those are the ones Fable was supposed to be for.

The honest summary is that this is a real upgrade with small print. The price cut is guaranteed. The performance gains are large but come mostly from the company’s own testing and a single independent evaluator in the first week. And the full 40% saving goes to people who let the new model run at its default rather than the one they chose for the old model.

Your Move

Have you switched yet? I’d like to know two things from anyone who has. Did the writing change show up in your own work, or only in the launch examples? And if you run it through the API, what did your cost per task actually do once you reset the effort level?

Share it in the responses. A week of real-world reports will say more about this model than any launch chart.

21 models, 1,050 head-to-head matches. Google’s cheap model beats its expensive one, second place costs $53 a month, and last place costs $250.

Somebody finally put 21 AI image models in a room and made them fight.

The format is the interesting part. Every model generates images for the same ten prompt groups, then the outputs go head to head across 1,050 pairwise matches, scored by an ensemble of six vision-language judges. No vibes, no cherry-picked demo gallery. Just the same prompts, the same scoring, and an Elo rating at the end, the way chess ranks players.

I went through the whole board. Here is who is actually winning, where the money is, and the two or three results that genuinely surprised me.

The podium

Gold goes to GPT Image 2.5 (OpenAI) on 1330. Silver to gpt-image-2 medium (OpenAI) on 1324. Bronze to MAI-Image 2.6 (Microsoft) on 1322.

OpenAI takes gold and silver. Microsoft takes bronze by two points.

Two points. That is the number to keep in your head for the rest of this, because the top of this board is not a ladder, it is a scrum. First to third is separated by eight Elo points in total. In chess terms, eight points means the better player wins about 51% of the time.

Then it gets tighter. Fourth is Grok Imagine Image 2.0 (xAI) on 1320. Fifth is nano-banana 2, listed as gemini-3.1-flash-image (Google), on 1316. Sixth is Muse-Image (Meta) on 1313. Seventh is gpt-image-1.5-high-fidelity (OpenAI) on 1311.

Seven models inside 19 points. Every major lab is represented. xAI showing up fourth with Grok Imagine Image 2.0 is the newcomer result of the board, and Meta landing sixth with Muse-Image means all five American frontier labs are now in the top seven.

OpenAI’s real advantage is depth, not the top spot

Holding first place is nice. Holding first, second and seventh is a different kind of statement.

The board publishes an average Elo per provider, and that is where the competition actually resolves. OpenAI with three models averages 1322. Microsoft with two averages 1316. Meta with one sits at 1313. Then Alibaba on 1305, ByteDance on 1303, xAI with two on 1302, Google with three on 1299, Black Forest Labs with two on 1288, Ideogram on 1286, NVIDIA on 1277, Recraft on 1271.

OpenAI’s average model scores 1322, which would place third on the overall board as a single entry. Three shots, three strong results. Anyone can get lucky with one model. Averaging near the podium across a whole lineup is harder to fake.

Google is the strangest story on the board

Google also fields three models. Google averages 1299, seventh place among providers, behind xAI and ByteDance.

That is not because Google’s best is weak. Its flash model sits fifth overall. The problem is the spread: fifth, ninth and twentieth. Three Google models, three very different tiers.

Now add the prices, because this is where I actually stopped and re-read the table. At 1,000 images a month on the rates the board lists, nano-banana 2 sits at rank five for $70 a month, while nano-banana-pro sits at rank nine for $150 a month.

The cheaper Google model ranks four places higher than the expensive one, at less than half the cost. If you had picked the pro model on the reasonable assumption that pro means better, you would be paying 2.1 times as much for a lower score.

I do not think that makes the pro model bad. Bigger models often win on things this benchmark does not isolate, and there may be jobs where it pulls ahead. But as a default purchase, on these numbers, it is very hard to justify.

Microsoft is the one nobody is talking about

Microsoft has two models on the board and both are quietly excellent: MAI-Image 2.6 at third with 1322, and mai-image-2.5 at eighth with 1309. Second-best provider average. No launch spectacle that I noticed, no discourse.

And then the price. mai-image-2.5 costs $50 a month at a thousand images. Eighth place on the board, cheaper than Google’s fifth-place model, cheaper than nearly everything above it.

If I were picking a workhorse today rather than a trophy, Microsoft’s pair is where I would start looking.

The price column changes everything

Here is the full cost list the board publishes at 1,000 images a month, cheapest first, with each model’s rank in brackets. grok-imagine-image (16) $20. gemini-2.5-flash-image (20) $39. nvidia/cosmos-3-super (19) $40. UNI-1.1 (17) $42. microsoft/mai-image-2.5 (8) $50. gpt-image-2 medium (2) $53. krea-2-large (13) $60. ideogram-4-fp8 (15) $60. flux-2-pro (18) $60. nano-banana 2 (5) $70. fal-ai/flux-2-max (12) $70. gpt-image-1.5-high-fidelity (7) $130. nano-banana-pro (9) $150. seedream-4.5 (14) $200. recraftv4_1_pro (21) $250.

Read the top and the bottom of that list together. Second place on the entire board costs $53 a month. Last place costs $250.

Recraft’s model finishes twenty-first of twenty-one and charges nearly five times what the silver medallist charges. Seedream 4.5 finishes fourteenth and charges $200. Meanwhile gpt-image-2 at $53 is beating nineteen models, several of which cost double or triple.

There is no relationship between price and rank on this board. None. If you assumed the expensive options were expensive for a reason, that assumption is not supported here.

That makes gpt-image-2 medium the value winner by a distance: rank two, $53. I would want to know what the medium setting gives up before committing a real workflow to it, but on the published numbers nothing else is close.

The cheap seats

The bottom of the board is not embarrassing, and this matters more than the ranking suggests.

The whole board spans 1330 down to 1271, which is 59 Elo points. Fifty-nine points is roughly a 58% win rate between the best and the worst model tested. The twenty-first-place model is not bad at making images. It is losing coin flips to the first-place model slightly more often than it wins them.

So the cheap end is genuinely interesting. grok-imagine-image at rank 16 costs $20 a month, the cheapest thing on the board by half, and it is xAI’s older model sitting five places below its own successor at fourth. gemini-2.5-flash-image at rank 20 costs $39, Google’s previous generation, still on the board. NVIDIA Cosmos 3 Super at rank 19 costs $40 and is the only open-weight model tested, which tells you plainly where frontier image generation currently lives.

If your work is high volume and quality-tolerant, thumbnails, mockups, internal decks, placeholder art, the $20 option loses head to head against the $53 option about 56% of the time. Put the other way, it wins roughly 44% of those matchups for 38% of the price. That is a real trade you might want to make.

The full 21

Rank one, GPT Image 2.5, OpenAI, 1330. Rank two, gpt-image-2 medium, OpenAI, 1324. Rank three, MAI-Image 2.6, Microsoft, 1322. Rank four, Grok Imagine Image 2.0, xAI, 1320. Rank five, nano-banana 2, Google, 1316. Rank six, Muse-Image, Meta, 1313. Rank seven, gpt-image-1.5-high-fidelity, OpenAI, 1311. Rank eight, mai-image-2.5, Microsoft, 1309. Rank nine, nano-banana-pro, Google, 1307. Rank ten, Qwen-Image 3.0 Pro, Alibaba, 1305. Rank eleven, Seedream 5.0 Pro, ByteDance, 1303.

Rank twelve, fal-ai/flux-2-max, Black Forest Labs, 1296. Rank thirteen, krea-2-large-text-to-image, 1292. Rank fourteen, seedream-4.5, ByteDance, 1289. Rank fifteen, ideogram-4-fp8, Ideogram, 1286. Rank sixteen, grok-imagine-image, xAI, 1284. Rank seventeen, UNI-1.1, 1282. Rank eighteen, flux-2-pro, Black Forest Labs, 1280. Rank nineteen, nvidia/cosmos-3-super, NVIDIA, 1277. Rank twenty, gemini-2.5-flash-image, Google, 1274. Rank twenty-one, recraftv4_1_pro, Recraft, 1271.

One honest caveat, because it affects how much weight to put on the exact rows. The board publishes 95% confidence intervals alongside these scores, and for a lot of adjacent pairs those intervals overlap, which means the gap between them sits inside the margin of error. Read this as tiers rather than a strict order. The top six is a real group. Fourteenth versus fifteenth is not a real distinction.

What I would actually pick

Best overall is GPT Image 2.5, if you can get it. It leads, and OpenAI’s depth across three models suggests the lead is not a fluke. Worth noting that the board lists no per-image price for it, so availability may be the deciding factor rather than quality.

Best value, and my actual default, is gpt-image-2 medium. Second on the board, $53 a month, beating nineteen models including several at triple the price. Best workhorse is microsoft/mai-image-2.5: eighth place, $50, from the second-strongest provider on the board, and badly underrated. Best cheap option is grok-imagine-image at $20, as long as you know you are trading a few percentage points of win rate for a 60% cost cut.

The one to think twice about is any expensive model in the bottom half. On this board, paying more does not buy a better rank.

Your Turn

The result I keep turning over is Google’s cheap model beating its own premium one by four places at half the price, because it means the naming conventions we all use as a proxy for quality are not tracking quality.

So: have you run a head-to-head yourself, and did the expensive model win? I am especially curious about anyone using the pro tier of anything and getting results the cheaper sibling could not match, because this board says that should be rare and my instinct says it depends entirely on the job.

Tell me what you are using and what it costs you. I will take real invoices over benchmarks.

Source: all rankings, Elo scores, provider averages and per-image costs in this piece come from the AI Image Generation Leaderboard published by Pixazo, as it stood in September 2026. The board scores 21 models across 10 prompt groups and 1,050 pairwise matches using a six-model vision-language judge ensemble. Monthly costs are the board’s figures at 1,000 images per month. Model names and version labels are reproduced as the board listed them and have not been independently verified.

I didn’t set out to run an experiment. I set out to plan a trip, settle an argument about a tax rule, and figure out why my sourdough kept collapsing. But somewhere around the third question, I noticed I’d started asking all three of my AI tools the same thing and quietly comparing their answers like a suspicious parent checking three kids’ stories. So I made it official. Over about ten days, I put the same 40 questions to Perplexity, ChatGPT’s search mode, and Google’s AI (the Gemini-powered results and the AI Mode). Same wording, same day, no cherry-picking. Here’s the honest version of what happened.

The setup (and why I stopped trusting one-shot answers)

A quick confession before the results: I used to take the first confident-sounding answer and run with it. That habit cost me an afternoon once when an AI told me a museum was open on a Monday. It was not. So my questions this time were deliberately mixed. Some were factual and checkable (opening hours, conversion rates, “what changed in this law”). Some were fuzzy and human (“help me plan three low-key days in Lisbon”). And a few were traps — questions with a false premise baked in, to see which tool would push back instead of playing along. That last category turned out to be the most revealing.

Perplexity: the one that shows its work

Perplexity was the tool I trusted fastest, and it took me a while to work out why. It’s the citations. Nearly every claim comes with a little numbered source you can click, and that changes how you read the answer — you skim, you spot the sentence that matters, you check where it came from. For my checkable questions (currency rates, whether a specific rule still applied in 2026, the actual boiling point of a technique I was arguing about), it was consistently the most sourceable. When I clicked through, the sources usually said what Perplexity claimed they said. That sounds like a low bar. It is not. It’s the whole game.

Where it fell down was warmth. Ask Perplexity to help you plan three unhurried days in Lisbon and you get something that reads like a very competent listicle assembled by a robot who has never been tired, never wanted a nap after lunch, never wanted to just sit somewhere and watch. Accurate, useful, a little joyless. Which, for research, is exactly the right trade.

ChatGPT Search: the one you actually want to talk to

ChatGPT’s search mode is the opposite personality. It’s the friend who read the article and is now telling you about it over coffee, with opinions. For the fuzzy, human questions it was easily my favorite. The Lisbon plan it gave me had a rhythm — it grouped things by neighborhood, it assumed I’d want to slow down, it suggested a bakery and then, unprompted, warned me the famous one has a queue and pointed me to a quieter one nearby. That’s the kind of judgment that makes an answer feel like advice instead of data.

The catch is that the same fluency that makes it pleasant makes it harder to audit. It cites sources, but less reflexively than Perplexity, and its confidence doesn’t visibly dip when it’s on shakier ground. Twice it gave me a smooth, plausible answer that turned out to be gently wrong — not wildly, just wrong enough that if I hadn’t cross-checked, I’d have repeated it at a dinner party. The better a tool sounds, the more you have to keep your guard up. That’s not a knock on ChatGPT specifically; it’s the tax you pay for eloquence.

Google AI: the one you’ll use whether you meant to or not

Here’s the thing about Google’s AI answers: you don’t choose them, they just show up. You search for something the way you’ve searched for twenty years, and now there’s a generated summary sitting on top. For quick, low-stakes facts — a conversion, a definition, when a shop closes — it’s genuinely convenient, and the speed is unmatched because you were already there. It won the “I need this in four seconds” category outright.

But the convenience is also the risk. Because it appears automatically and looks like part of Google itself, it borrows Google’s authority even when it shouldn’t. A couple of times the summary blended two different sources into one confident sentence that neither source actually supported. And for anything nuanced, it flattened. Ask it a fuzzy planning question and you get a tidy paragraph that’s fine but forgettable. It’s the tool most likely to be right about small things and most likely to be quietly trusted when it’s wrong.

The trap questions, where it got interesting

My favorite part of this whole thing was the false-premise questions. I asked things that assumed something untrue — phrased as if a fact were settled when it wasn’t, or asking “why” about something that never happened. This is where the personalities really separated. Perplexity was the most likely to quietly correct me, because its instinct is to go find sources, and the sources didn’t back up my premise, so it said so. ChatGPT sometimes caught it and sometimes cheerfully answered the question I’d accidentally implied, building a confident explanation for a thing that isn’t real. Google’s summary was the most likely to just roll with the false premise, because it’s optimized to answer fast, not to argue with you.

If you take one thing from this, take that. The most important quality in an answer engine isn’t how smart it sounds when you’re right. It’s whether it pushes back when you’re wrong. A tool that won’t correct your premise will happily walk you off a cliff with perfect grammar.

So which one should you use?

Annoyingly, the honest answer is all three, for different jobs. When I actually need to be right — something I’m going to publish, cite, or make a decision on — I reach for Perplexity, because the citations let me verify instead of trust. When I want to think something through, plan something human, or get advice with a bit of judgment in it, ChatGPT’s search mode is the one I enjoy and the one that gives me something worth reacting to. And Google’s AI is the one I use without deciding to, for the tiny throwaway facts where being 95% right in four seconds is genuinely enough.

What surprised me most is that after ten days, my takeaway wasn’t about which tool won. It was about how much my own habits had to change. These things are extraordinary at sounding right, and that’s precisely the problem. The skill that matters now isn’t finding an answer — answers are everywhere, instant and free. It’s keeping a little suspicion alive. Clicking the source. Noticing when a reply is a bit too smooth. Asking the follow-up. The tools got better; the part where I stay awake at the wheel is still on me.

Your Turn

I’m curious which one you actually reach for first — and whether you trust it more than you should. Have you ever repeated something an AI told you confidently, only to find out later it was wrong? Or caught one of them quietly making things up? Tell me in the comments. I read all of them, and the best trap question I’ve seen so far came from a reader, so bring your sneakiest one.

I spent a month running the same projects through all three. Here’s where each one actually fits — and where the hype quietly falls apart.

I almost didn’t write this one. Another AI image tool comparison is exactly the kind of thing the internet already has too much of, and I could feel myself about to add to the pile. But after a month of pushing the same real projects through Midjourney, Adobe Firefly, and Ideogram — not cute test prompts, but actual client work and personal pieces I cared about — I ended up with opinions strong enough that staying quiet felt a little dishonest.

So this is the honest version. Not a spec sheet you could pull from three pricing pages, but where each tool genuinely earned its place in how I work, and where it quietly let me down when it actually mattered.

For context: I’ve been generating images with these tools more or less daily for over a year now, across enough styles and dead ends to have lost most of my early enthusiasm and replaced it with something more useful — patterns.

The setup: same prompts, real stakes

I gave all three the same five jobs: a blog header with text baked into the image, a set of moody product shots, a few character illustrations that needed to stay consistent, some abstract background textures, and a round of logo concept exploration. I ran every prompt at least three times in each tool. Anyone who claims they nailed the perfect image on the first attempt is either unusually lucky or gently lying to you.

What surprised me wasn’t which tool ‘won.’ It was how little they actually overlap. I went in expecting three flavors of the same thing. I came out convinced they’re three different instruments, each best at something the others noticeably fumble.

Midjourney: still the one with taste

Midjourney is still the tool that makes me say ‘oh, that’s nice’ out loud before I’ve even thought about it. Its default sense of lighting, composition, and that faintly cinematic mood is ahead of everyone else. For the moody product shots and the abstract textures, nothing else came close. I’d hand it a half-formed idea and it would return something more interesting than what I’d pictured.

The catch is control. Midjourney is dazzling at vibe and stubborn about specifics. When I needed a product in an exact spot, or a character to survive across a whole series without subtly changing faces, I spent more time wrestling it than I’d like to admit. It’s the wildly gifted artist who’s very hard to art-direct — brilliant when you let it roam, exhausting when you need a precise result.

And text? Just don’t. It has improved, and I want to be fair about that, but asking Midjourney to render a clean headline still feels like asking a painter to do your taxes. Technically possible. Rarely worth it.

If your work is editorial, conceptual, or mood-driven — book covers, atmosphere pieces, anything where ‘how does it feel’ beats ‘is the logo in the right corner’ — Midjourney still deserves the first slot in your rotation. I just stopped expecting it to also be my layout tool.

Ideogram: the one that can actually spell

This is where Ideogram quietly stole the whole project. The blog header that needed readable text inside the image — the thing every generator has historically mangled into confident-looking alien gibberish — Ideogram handled on the second try. Legible, sensibly placed, spaced like someone who has actually seen typography before.

For anything involving words — posters, mockups with real copy, social graphics, early logo concepts — Ideogram became my default without me really deciding it. It isn’t as painterly as Midjourney, and its purely ‘artistic’ outputs can feel a little generic next to it. But when the words on the image matter, it isn’t a close contest.

A reader pushed back on an earlier piece of mine about whether text rendering was really such a big deal. After this month, my answer is an unqualified yes. When I counted how many images I’d quietly thrown away over the past year purely because the text was unusable, the number was higher than I wanted to believe.

The other quiet win was iteration speed on text-heavy work. Because I wasn’t regenerating fifteen times hoping the letters would finally cooperate, a job that used to eat an afternoon collapsed into about twenty minutes. That’s not a flashy benefit, but it’s the kind that changes whether you actually enjoy the work.

Firefly: the one you can actually use at work

Adobe Firefly is the least thrilling of the three and, depending on what you do for a living, the most important. Its images rarely made me gasp. What it offers instead is peace of mind: it’s trained on licensed and public-domain content, which means I can hand the result to a client without that low background hum of ‘wait, where did this training data actually come from?’

For commercial work, that isn’t a footnote. It’s most of the game. Firefly also lives inside Photoshop, so generative fill and the small fixes that eat your afternoon happen right where I’m already working, instead of in a fourth browser tab I’ll forget to close. The integration is the feature, more than any single image it produces.

Would I reach for it during a late-night personal art experiment? Rarely. Would I use it for anything a client is paying real money for? Increasingly, every time. There’s also something underrated about predictability — Firefly rarely surprises me to the upside, but it almost never surprises me to the downside either, and on a deadline that boring reliability is worth more than the occasional brilliant accident.

So which one should you actually use?

Here’s the unsatisfying but honest answer: probably more than one. Somewhere in week two I stopped treating these as rivals and started treating them as a small team with different specialties, and everything got easier.

If you care most about mood and aesthetics, Midjourney. If your image needs words in it, Ideogram, full stop. If you’re doing commercial work where licensing and integration matter, Firefly. The real mistake — and I think it’s the source of most of the frustration you read online — is picking one tool and then resenting it for not being good at the two things it was never going to be good at.

The bigger shift for me was realizing that ‘what’s the best AI image tool’ is just the wrong question. It’s a bit like asking whether a pencil is better than a camera. The better question is what you’re trying to make today, and which tool will get out of your way the fastest while you make it.

If anything, running them side by side made me more patient with each one. Once I knew what each was for, the failures stopped feeling like the tool being broken and started feeling like me reaching for the wrong one. That reframe alone has saved me more frustration than any prompt trick I’ve picked up.

Your turn

I already know my experience won’t match yours, and I’m genuinely curious where they split. Which of these three has earned a permanent place in your workflow, and which one did you quietly abandon once the novelty wore off? And if you’ve landed on a combination that works, I’d honestly like to steal it — tell me about it in the responses.

The tools got shockingly good. The hard part just moved somewhere I didn’t expect.

About a year ago, I started keeping a folder on my desktop called “AI tests.” Every time a new model dropped or a tool promised to change everything, I’d throw a real project at it and save the results. The folder is now embarrassingly large. Going back through it recently, I noticed something I didn’t expect: the difference between my early work and my recent work has almost nothing to do with the tools. The tools were always capable. I just didn’t know what to ask them for.

That realization has reshaped how I think about creating with AI. We spend so much energy arguing about which model is best, which prompt structure wins, which app has the slickest interface. But after a year of testing, the thing that actually separated good output from forgettable output wasn’t technical at all. It was taste.

The tools stopped being the bottleneck

Think about what’s happened in the last eighteen months. Image models that once produced melted hands now render photos you’d swear were real. Writing tools went from clumsy to genuinely fluent. Video, music, code — the floor of “good enough” rose so fast that most of us are still recalibrating.

When everyone has access to the same capable tools, capability stops being the differentiator. I can generate fifty images in the time it used to take to sketch one. But fifty mediocre images aren’t worth a single good one, and the model has no idea which is which. That judgment is mine to make. The bottleneck moved from “can I make this?” to “can I tell what’s actually worth making?”

What taste actually means here

Taste is a slippery word, so let me be concrete about what I mean. I don’t mean some innate gift you either have or don’t. I mean the accumulated ability to look at two options and know which one is better — and, crucially, to know why.

When I ask an image model for a “minimalist book cover,” I now have opinions about negative space, about where the eye lands first, about whether the type feels expensive or cheap. I didn’t have those opinions a year ago. I developed them by looking at thousands of outputs and slowly learning to articulate what was wrong. The model gave me the reps. The taste came from paying attention during those reps.

This is the part people miss. AI doesn’t remove the need for judgment — it raises the premium on it. The more output you can generate, the more your ability to curate becomes the scarce resource.

And here’s the uncomfortable corollary: if curation is the scarce resource, then most of the value you add never shows up in the final piece. Nobody sees the forty images you rejected or the eleven headlines you killed. They only see the one you kept. That invisible labor — the deciding, the discarding, the quiet “no, not that” — is exactly the part the model can’t do for you, and it’s the part that’s easiest to skip when you’re tired or rushed. I’ve learned to treat that fatigue as a warning sign rather than a green light. The moment I stop caring which version I ship is the moment the work starts looking like everyone else’s.

The curation muscle

Here’s a small experiment that changed how I work. For a week, I forced myself to generate three versions of everything and then write one sentence explaining why I picked the one I picked. “This headline is better because it makes a promise instead of describing a topic.” “This image works because the subject is off-center and it feels less like stock photography.”

It was tedious. It was also the most useful week I’ve had with these tools. Forcing myself to articulate the why turned a vague gut reaction into something I could repeat. The next time, I could skip straight to the better version because I understood the principle underneath it.

Most people use AI the opposite way. They generate, glance, accept the first thing that’s “fine,” and move on. That’s the workflow that produces the flood of generic content we’re all starting to recognize and tune out. The generic isn’t the model’s fault. It’s what happens when nobody brings judgment to the table.

I think this is why the “articulate the why” habit matters so much. It turns taste from something mystical into something teachable. When I can name the reason one option beats another, I can apply that reason a hundred times without re-deriving it. I can also hand it to someone else, or to the model itself in the next prompt. Taste that stays trapped as a feeling helps only you, in the moment. Taste you can put into words becomes a system — and systems are what let you stay consistent when you’re producing at the speed these tools now allow.

Why this is actually good news

If you’ve been anxious about AI flattening creative work, I think this is genuinely reassuring. The skill that matters most now is one that compounds with experience and can’t be downloaded. A model can generate a thousand options instantly. It cannot care which one is right. It has no point of view, no sense of who the work is for, no memory of the specific human you’re trying to reach.

That’s the gap where real creators live now. Not in the mechanical production — the model handles that — but in the direction. Knowing what to make. Knowing when something is done. Knowing when “technically impressive” is actually worse than “simple and clear.” These are taste decisions, and they’re getting more valuable, not less.

How I’d build it if I were starting today

If I were new to all this, I wouldn’t start by collecting tools. I’d start by collecting examples of work I admire and trying to figure out, in writing, what makes them good. I’d generate a lot and throw most of it away on purpose. I’d treat every rejected output as a small lesson rather than a failure.

I’d also resist the urge to chase every new release. The tool you already know well, used with real judgment, will beat the shiny new one used carelessly almost every time. (I’ve fallen for the shiny-new-thing trap more times than I’d like to admit — there’s always one more platform promising to be the last one you’ll ever need.) The capability is rarely the limiting factor anymore. You are. And that’s a far more interesting problem to work on.

Your Turn

I’m curious whether this matches your experience. If you’ve been creating with AI for a while, do you feel the bottleneck has moved for you too — away from the tools and toward your own judgment? Or do you think I’m overstating it, and the tools still have a long way to go?

Drop a comment and tell me about a moment when your taste, not the tool, made the difference. I read every response, and the best conversations here always seem to start in the comments.

Three months ago, I realized I couldn’t write a single paragraph without opening an AI chat window first.

Not because I needed help — at least, that’s what I told myself. It was more of a reflex, like checking your phone the moment you wake up. I’d have an idea, sit down at my desk, and before I’d written a single sentence, I’d already pasted my half-formed thought into Claude and asked it to help me “find the angle.”

I’m not saying that’s wrong. But when I noticed I was doing it automatically, without thinking, I got a little unsettled. Not because AI was making my work worse — objectively, it wasn’t. But because I wasn’t sure anymore whether I still knew how to start without it.

So I decided to find out.

The Ground Rules

No AI writing tools for seven days. No Claude, no ChatGPT, no Gemini. No Perplexity for research. No AI grammar checkers, no “polish this paragraph” shortcuts. Just me, a blank document, and whatever I still remembered about how I used to work.

I didn’t tell anyone I was doing this. It felt embarrassingly small as an experiment — like announcing you’re taking a break from Instagram. But the results were strange enough that I think they’re worth sharing.

Days 1 and 2: Slower Than I Expected

The first day was uncomfortable in a way I hadn’t anticipated. Not frustrating, exactly. More like trying to write with your non-dominant hand. Everything felt effortful in a way it hadn’t for a while.

I noticed two things right away. First, I was staring at the blank page longer. A lot longer. What used to be a 10-minute warm-up — paste rough notes, get a framing angle, start writing — became 40 minutes of actually sitting with the blankness, not knowing what I wanted to say.

The second thing: my internal editor was louder than I remembered. Every sentence I wrote, I was already judging it. I kept expecting some outside voice to tell me whether I was on the right track, and when that voice didn’t come, the doubt filled the space instead.

By the end of Day 2, I’d written about 600 words. It felt like pulling teeth. But it also felt unmistakably mine in a way I’d half-forgotten.

Days 3 and 4: Something Started Clicking

By the middle of the week, something shifted. I’m not sure exactly what triggered it — maybe just adaptation, maybe boredom becoming focus. The blankness stopped feeling threatening and started feeling like space.

I started doing something I hadn’t done in months: going for walks just to think. Not productive walks with a podcast. Just walks — moving around the neighborhood, not consuming anything, letting my brain run loose on whatever I was working through. I’d come back and write for an hour, and what came out was rougher at the sentence level, but more alive in some way I found hard to name at the time.

I also noticed I was making stranger choices. Not necessarily better ones. But more idiosyncratic — more specific to how I actually see things. When I’m working alongside AI, there’s a subtle gravitational pull toward the sensible interpretation, the expected framing. Not because AI imposes it, but because having any outside perspective tends to nudge you toward what communicates clearly rather than what surprises.

Without that nudge, I found myself going places I wouldn’t normally go. Some of it was dead ends. But some of it was the most interesting writing I’d done in a while.

Days 5–7: What Actually Got Better (And What Didn’t)

By the end of the week, I’d found a rhythm that was slower but more intentional than what I’d been doing before. Some things genuinely improved. Some things were just worse.

What got better: voice. The specific texture of how I phrase things, the rhythms I fall into naturally — it was more pronounced when I wasn’t calibrating against anything external. I was just writing the way I actually write.

My ideas also landed on the page less finished, which sounds like a disadvantage but wasn’t. Ideas that aren’t solved yet leave more room for the writing to figure them out. When I’ve already “resolved” a concept in an AI conversation before I start writing, the article is sometimes little more than transcription. This week, I was genuinely discovering things mid-sentence.

What didn’t get better: research. Without Perplexity, I lost entire mornings to browser rabbit holes. I’d set out to verify a single data point and surface 45 minutes later, slightly dazed, having read four tangential articles and bookmarked a dozen more. AI hadn’t been making me lazy with research — it had been protecting my attention.

Also: rough draft quality was noticeably lower. More filler sentences, more vague assertions, more places where I could feel myself gesturing at an idea rather than actually nailing it. I’d gotten used to AI as a mirror that showed me where my thinking was incomplete. Without the mirror, the incompleteness slipped through.

What I Brought Back, and What I Left Behind?

When the week ended, I didn’t just snap back to my old habits. Spending seven days noticing them had made them impossible to unsee.

What I brought back: AI for research and for review. Perplexity stays — it’s too useful for protecting attention to give up. So does using Claude as a thinking partner after I’ve written a first draft, not before. The mirror function is genuinely valuable. The scaffolding function, I’ve decided, isn’t for me.

What I left behind: the reflexive reach. The opening-an-AI-tab-before-I’ve-thought-for-five-minutes habit. The asking-for-an-angle before I’ve sat with the blankness long enough to find my own.

I’ve started giving myself 20 minutes of no-AI time at the start of every creative session. Just me and the blank page, getting uncomfortable together. Then, after I’ve written something — even something rough and imperfect — I’ll bring in the tools.

The difference has been real. Not dramatically better on any measurable metric. But more distinctly mine. Which, for what I’m trying to build here, matters more than I’d been letting myself admit.

The Thing Nobody Warns You About AI Dependency

There’s a lot of conversation about whether AI makes creative work better or worse. I think that’s almost beside the point.

The more interesting question is whether you’ve built your creative process around the tools in a way that makes them load-bearing. Whether, if you took them away, you’d be left with something solid — or discover that what felt like muscle was actually scaffolding.

I use AI tools every day and I’m not going to stop. They’re genuinely powerful, and pretending otherwise would be dishonest. But this week reminded me that the most important creative skill I have isn’t prompting or workflow optimization — it’s the capacity to sit with a blank page and generate something from nothing. That capacity is use-it-or-lose-it. And I’d been quietly losing it.

I’d recommend this experiment to anyone who uses AI heavily in their creative work. Not as a judgment — go back to the tools after, absolutely — but as a calibration. You might be surprised by what you find on the other side of a week without them.

Your Turn

I want to ask you something honestly: do you still know how to start without AI? Not as a rhetorical challenge — I’m genuinely asking, because I’d half-forgotten, and that surprised me more than anything else this week.

If you’ve done a similar experiment — voluntary or forced — I’d love to hear what you found. And if the idea of a week without AI tools feels genuinely alarming, that might be the most useful data point of all.

Drop your thoughts in the comments. This is one I’m still sitting with.

I remember mine clearly. I had a dozen AI tools bookmarked, three subscription payments going out each month, and a workflow that had somehow become more complicated than the one I started with. I was spending more time managing AI outputs than I was actually creating.

So I did what any sensible person would do: I quit almost all of it.

Then, slowly, I started over — this time paying attention to what was actually helping versus what I was using because it felt like I should be using it.

What follows is my honest, opinionated take on what stuck and why. Not a sponsored listicle. Not a breathless review of every new release. Just the things that actually made it into my daily workflow after more than a year of experimentation.

The First Mistake: Treating AI Like a Vending Machine

The biggest mistake I see creatives make — and I made it too — is treating AI like a vending machine. You put in a prompt, you get out content, you move on.

The problem is that creativity doesn’t actually work like that. The value isn’t in the output. It’s in the process of thinking, iterating, reconsidering, and finding the unexpected thing you weren’t looking for.

When I started using AI as a vending machine, I got faster output but worse work. My writing lost its texture. My design concepts felt generic. I was producing more, but I was less proud of any of it.

The tools that I kept are the ones that helped me think better, not just work faster. That distinction matters more than anything else in this conversation.

What Actually Stayed in My Workflow?

The first thing that genuinely changed how I work is using AI as a thinking partner, not a content generator. I’ll put a half-formed idea into a conversation with an AI assistant — I’ve been using Claude for this lately — and then push back on everything it says until I figure out what I actually believe. It’s like having an infinitely patient colleague who will take your weird 11pm idea seriously and help you poke holes in it.

The difference from my early vending-machine approach is that I’m still doing the creative work. The AI just makes it easier to externalize my thinking so I can evaluate it.

The second keeper is Midjourney — but not the way most people use it. I almost never use Midjourney outputs as finished images. Instead, I use it for direction. Before starting a visual project, I’ll generate 20 or 30 rough concepts to understand what aesthetic territory I want to be in. It’s like a mood board I built in 20 minutes instead of 3 hours. Then those become a jumping-off point for actual work.

The key shift was accepting that AI output is a starting point, not an ending point. When I treat generated images as reference material rather than finished work, I get the speed benefit without the homogenization problem.

The third tool is less glamorous: Notion AI for administrative overhead. Meeting notes. Draft outlines. Summarizing research I’ve already gathered. Generating first-draft briefs that I then rewrite entirely. It saves me roughly an hour a day on cognitive overhead work that doesn’t require my best thinking — and that gives me more energy for the stuff that does.

What I Dropped (And Why)

I tried AI image upscalers for about two months. They’re technically impressive, but every output felt slightly wrong in a way I couldn’t fully articulate. Too smooth. Too confident in its hallucinations about what should be in a shadow.

I dropped automated social media caption generators after the first week. The captions were fine. That’s the problem — “fine” is the enemy of a distinctive voice.

I stopped using any tool that tried to fully automate the writing of anything I’d put my name on. Not because AI writing is bad. But because the process of writing is the creative act, and automating it away means you’ve automated away the thing that made it worth doing.

The Pattern I Kept Seeing

Looking back at what stayed versus what didn’t, there’s a clear pattern. The tools that remained are the ones that amplify the parts of creativity I find hard — the ideation, the stuck moments, the administrative overhead. The ones that left are the ones that tried to replace the parts I find meaningful.

That might sound obvious. But I think a lot of us, when we’re first integrating AI into our work, don’t have a clear enough picture of which parts we actually find meaningful. We assume we want to go faster everywhere. We don’t.

A Question Worth Asking Yourself

Before you try a new AI tool, ask yourself: “Is this for the part of my work I want to speed up, or the part that’s actually the point?”

It’s a surprisingly clarifying question. Most of the time, I find the things I most want to automate are secretly where the good work lives.

Administrative and logistical tasks — absolutely, let AI handle as much of that as possible. The ideation, the refinement, the genuine creative choices? That’s the part worth protecting.

After more than a year of daily AI experimentation, I’m genuinely enthusiastic about its role in creative work. But I’m enthusiastic about specific applications, not the vague promise that “AI will change everything.” It will — but only if you’re selective about what you hand over and what you keep.

The best version of AI in your creative workflow is one where you’re doing more of what you love, not less.

Your Turn

What’s one AI tool you tried and dropped? And what made you drop it? I’m curious whether other creatives are finding the same patterns — or completely different ones. Drop it in the comments.

I’ve been at this long enough now to have a graveyard of apps, subscriptions, and bookmarks I’ll never open again. Tools that seemed revolutionary for about two weeks before quietly disappearing from my routine. Platforms that promised to transform my creative process but ended up just adding friction.

And I don’t think that’s a failure. I think it’s the natural arc of how we actually learn what works.

So today I want to do something a little different. Instead of reviewing the newest shiny thing, I want to be honest about what my actual day-to-day AI toolkit looks like — the tools I genuinely open every morning, and the ones I convinced myself were essential before realizing they weren’t.

Let me get this out of the way, because I think it’s more useful than the list of winners.

Over the past several months, I’ve tested more AI tools than I can count. Most of them had genuinely impressive demos. Some of them had features I still wish the tools I actually use had. But they didn’t make the cut for a few consistent reasons.

They were too slow in practice. A tool that produces brilliant output but takes three minutes per generation kills creative flow faster than anything else. Speed matters — not just render speed, but how fast you can go from an idea to something you can actually react to.

They required too much setup. The best AI tools disappear into your workflow. The worst ones feel like a part-time job to configure before you can even start. If I’m spending twenty minutes per session adjusting settings, the tool is working against me, not for me.

They optimized for demos, not for actual use. I can’t tell you how many tools looked extraordinary in a tutorial video and then produced mediocre results on anything other than the exact type of prompt the creator used. Real creative work is messy. Your tools need to handle messy.

They tried to do everything. The AI do-everything platforms are almost always worse at each individual thing than specialized tools. Trying to replace your entire workflow with a single platform sounds appealing in theory. In practice, it usually means getting mediocre results across the board.

What Actually Survived?

1. An LLM as a thinking partner, not a writing machine

Here’s what I use every day — not because someone paid me to say it, but because my actual morning routine depends on these things and I would notice immediately if any of them disappeared.

Most people use large language models as writing tools: give it a prompt, get text out, edit it, call it done. That’s not how I use them anymore.

I use mine as a thinking partner. I give it half-formed ideas, contradictory thoughts, things I’m not sure how to feel about yet — and I have conversations with it. It pushes back when my logic has gaps. It asks clarifying questions. It helps me figure out what I actually think before I start writing.

For creative work, this has been more valuable than any amount of text generation. The writing is still mine. But having a tool that helps me get to the point where I have something worth writing — that’s a completely different and more useful thing.

2. An image generation tool for visual thinking, not final output

I want to stay deliberately vague on which one, because the landscape shifts fast and what works best depends heavily on your aesthetic sensibility and workflow.

What I can say is this: I stopped using image AI to make finished work and started using it to think visually before committing to a direction. Rough visual exploration takes twenty minutes with AI. Refined execution of a clear creative direction takes two hours without it. The math is obvious.

The shift that made this work for me was treating AI-generated images as sketches, not products. As soon as I stopped caring whether the output was good, I started getting enormous value from it.

3. A voice capture tool for ideas that happen away from a desk

This one surprised me the most, and it’s also the least glamorous item on this list.

I started using voice-to-text transcription — one of the newer AI-powered ones on my phone — to capture ideas while walking, cooking, or doing anything that isn’t sitting at a desk. The raw transcripts are messy. I don’t polish them. I just dump them into a folder and revisit them when I sit down to write.

About a third of my best ideas in recent months came from those voice notes. Ideas I would have lost completely if I’d waited until I was back at a keyboard. If you’re a creative person who does your best thinking away from a screen, this category is worth exploring.

4. A writing refinement tool for the mechanical stuff

Not for drafts. Not for first ideas. But for things like: does this sentence make sense? Is this section too long? Am I repeating myself in a way that’s going to bore a reader?

I use writing AI tools the same way I’d use a patient editor who never gets tired of reading my drafts. The ideas are mine. The decisions are mine. But having something that can flag when a paragraph is doing too much work, or when I’m being unclear without realizing it — that’s genuinely useful in a way that doesn’t compromise the work.

5. An automation layer that handles everything I shouldn’t be doing manually

This is the boring one, but it’s probably where I save the most time overall. I use AI-powered automation to handle the administrative tasks that surround creative work: organizing research, reformatting files, summarizing long documents I don’t have time to fully read before a deadline.

None of this is creative. None of it is interesting. But it’s a non-trivial part of what creative work actually requires, and getting AI to handle it means I have more time and energy for the parts only I can do.

The Pattern Worth Noticing

Looking at this list, something becomes clear: every tool I use daily is doing something distinct. They don’t overlap or compete with each other. Each one covers a different phase of the creative process — ideation, visual exploration, idea capture, refinement, administration.

The tools I abandoned were almost always either duplicating something I already had, or trying to compress the entire creative process into a single button. Neither approach works for how I actually make things.

What This Means for You

If you’re in the stage of testing every new tool that launches — and I’ve been there, very recently — here’s what I’d actually suggest: stop optimizing for comprehensiveness and start optimizing for repetition.

The question isn’t what AI tool is best. It’s what do I actually use every week, consistently, without having to think about it. That’s your real toolkit. Everything else is interesting — some of it genuinely impressive — but impressive and essential aren’t the same thing.

The tools that will change your work long-term are the ones you eventually stop thinking about as tools. They just become part of how you work. That quiet integration is what you’re actually looking for.

Your Turn

I’m genuinely curious: what’s in your actual everyday AI toolkit — not the stuff you tell people you use, but the things you’d immediately notice if they disappeared tomorrow? Drop it in the comments. I read every single one.

I’ll be honest: until about a month ago, I mostly tuned out whenever someone started talking about “AI agents.”

It sounded like something for engineers. Something that involved APIs and code and infrastructure decisions I wasn’t qualified to make. I’d hear terms like “multi-agent orchestration” and my brain would quietly file it under interesting but not for me — right alongside “Kubernetes” and “distributed systems.”

Then something shifted.

I was using an AI tool to help me brainstorm a concept for a short creative project, and instead of just giving me a list of ideas, it started doing things on its own. It searched for references. It pulled together thematic notes. It drafted a one-page outline without me asking. And then it asked me: “Would you like me to look for similar projects that got traction recently?”

I sat back and thought: oh. This is different.

That’s when I started paying real attention to what AI agents actually are — and why I think they’re going to matter a lot to people who make things for a living.

So What Even Is an AI Agent?

Let me skip the jargon and give you the version that made sense to me.

A regular AI tool — the kind most of us have been using for the past couple of years — responds to what you ask. You type a prompt, it gives you an output. You’re the one doing all the orchestrating. You decide what to ask next, where to take things, what to do with the result.

An AI agent is different because it can take sequences of actions toward a goal, often without you having to specify every step. You give it a direction, and it figures out a path. It might search the web, write something, revise it, check a fact, and then deliver you something useful — all in one go.

Think of it like the difference between a paintbrush and a skilled collaborator who can pick up a paintbrush and run with a concept you sketched on a napkin.

Does that mean it replaces you? No. I’ve seen what happens when agents run without enough direction and the output is… not great. But as a tool that works with you, the potential is genuinely interesting.

Where This Gets Interesting for Creatives

Here’s my honest take: the first wave of AI tools were mostly useful for the mechanical parts of creative work. Drafting first versions. Generating reference images. Summarizing research. Genuinely helpful — but the creative judgment was entirely on us.

Agents start to touch something different. They can handle the connective tissue of a project — the stuff between the creative leaps that’s still mentally taxing even if it isn’t especially creative.

Let me give you some concrete examples from my own work lately:

When I’m developing a new article, I don’t just write. I research, find examples, check whether my claims hold up. An agent can take a rough angle and do a lot of that background work while I’m thinking about structure and voice. By the time I sit down to actually write, I’m better prepared.

When I’m exploring a visual concept, I don’t start from a blank canvas — I start with references, moods, color palettes, compositional ideas. An agent can assemble a lot of that for me if I give it the right direction. Not the creative vision, but the supporting material.

And for longer projects — video, campaigns, long-form writing — the task management overhead can be enormous. Keeping track of what’s done, what needs revision, what’s waiting on something else. Agents are starting to help here too.

None of this means the creative work is getting done for you. It means the time you spend on things that aren’t the creative work starts to shrink.

The Part That Makes Me Cautious

I want to be careful not to oversell this, because I’ve been burned by AI hype before.

The agent tools available right now are still fairly limited in real-world creative use. They make mistakes. They sometimes go off in directions you didn’t intend. The “autonomous” part of autonomous agents is still pretty supervised in practice — you need to watch them closely.

There’s also a subtler risk I think creatives in particular need to watch for: the more an agent handles the research and preparation phase of your work, the more your creative output risks being shaped by what the agent finds — which is shaped by what already exists, is indexed, and is retrievable online. The weird, specific, idiosyncratic stuff that makes a creative perspective original? An agent can’t find that for you. That’s still entirely yours.

So I think of agents as a way to handle volume and logistics, not a way to replace the curiosity and lived experience that make your work worth reading or seeing.

What I’m Actually Doing Differently

Since I started treating agent-style tools as part of my workflow, a few things have changed practically.

I front-load my creative process more. I give the agent a direction early and let it gather material while I do other things, rather than sitting at a blank page waiting for inspiration to arrive.

I’ve started thinking about what I call the “handoff point” — the moment where I take over from what the agent has assembled and apply my actual judgment. That handoff is everything. Before it, I’m setting direction. After it, I’m being a creative.

I’ve also gotten more deliberate about what I’m not delegating. My voice, my conclusions, the specific examples from my own experience — those stay with me. An agent can draft, but it can’t observe my life. It can’t replicate the particular angle I bring to a subject, the things I’ve noticed and cared about. That has to be mine.

One unexpected shift: I’ve become more opinionated, not less. When the structural work gets easier, you have more mental space to actually commit to a perspective. I think that’s a good thing.

Where Things Are Headed (My Guess)

I think within a year or two, “using AI agents” will feel as unremarkable as “using the internet to research” does now. It’ll be a standard part of how creative work gets done — not a special capability, just a tool in the kit.

What that means is that the premium will increasingly be on things agents genuinely can’t replicate: original perspective, genuine lived experience, aesthetic judgment, the willingness to take creative risks that don’t have a precedent in training data.

That’s both exciting and a little pressure-inducing, honestly. The bar for what makes creative work valuable is going to shift. Work that’s technically proficient but generic won’t stand out the way it once might have. Work that’s genuinely surprising, personal, or risky will.

I’d rather know that’s coming and think about it now than be caught off guard later.

The creatives who figure out how to use these tools without losing what makes their voice distinctly theirs are going to be in a really interesting position. That’s what I’m trying to work out, one project at a time.

Your Turn

I’m curious how you’re thinking about this. Have you started using any of the more agent-style AI tools in your own creative process? Where do you see them fitting — or not fitting at all?

I’m especially curious from people working in visual fields — illustration, video, design — because I have a feeling the implications look quite different there than they do for writers. The relationship between the artist’s eye and an agent’s output seems like it would create a whole different set of questions.

Drop a comment below. I read all of them, and conversations like this are genuinely where I learn the most.