Every one of these models can spit out a photorealistic Mona Lisa from a text prompt in seconds. So a researcher took that away and asked a harder question: can they draw one? In July 2026, TryAI built a "canvas arena" — a blank white canvas, a set of colored-pencil tools, and four frontier vision models (GPT-5.6 Sol, Claude Fable 5, Grok 4.5, Gemini 3.6 Flash) told to reproduce the Mona Lisa stroke by stroke. The models could set a color, tip width, and pressure, lay down strokes, smudge to blend, erase, view the target, view their own canvas, and keep going. Every stroke, dollar, and score was tracked, and the results are more interesting than the pictures. This guide walks the whole experiment: the setup, the SSIM scores (Gemini 0.337, GPT-5.6 Sol 0.325, Claude Fable 5 0.286, Grok 4.5 0.151 on the Mona Lisa), the 20x cost gap that did not buy better output, and the four genuinely different strategies the models improvised from one identical toolset. Then it draws out the lesson that actually matters for anyone building content with these models: generating an image and executing one are different skills, and the execution one is where they are still uneven, expensive, and — critically — unable to tell when to stop. Every scored run peaked mid-way and then got worse, because the models kept editing past their best frame with no taste to say "done." That failure is the whole point, and it is exactly why raw model capability is not the same thing as finished, on-brand, shippable output.
Ask GPT-5.6, Claude, Gemini, or Grok to "show me the Mona Lisa" and you get a clean, photorealistic image in seconds. That is generation, and it is close to a solved problem. So in July 2026 the team at TryAI removed generation from the table and asked a much harder question: can these models draw? They built what they called a canvas arena — a blank white canvas and a set of colored-pencil tools — and told four frontier vision models to reproduce famous images, the Mona Lisa and Van Gogh's Starry Night among them, one stroke at a time. No image generator, no fill bucket, no shortcut. Just a pencil, an eraser, the ability to look at the target and at their own work, and instructions to keep going until it looked right.
The point of that constraint is the whole point of this guide. Generating an image is a single feed-forward act: the model produces pixels and it is done. Drawing one is an agentic loop — plan a stroke, place it, look at the result, decide what is wrong, correct it, repeat, and eventually judge that it is finished. That loop is where visual reasoning, tool use, and self-critique all get tested at once, and it is a far better probe of what a model can actually do with its "eyes" than a benchmark score. The four models tested were GPT-5.6 Sol, Claude Fable 5, Grok 4.5, and Gemini 3.6 Flash — the same frontier tier covered in the 2026 image and video model review. What came back was four wildly different attempts and one shared, revealing failure.
Each model got the same seven-ish tools and nothing more. It could plan on a private scratchpad, view the target image, and view its own canvas. To make a mark it set a color, set a brush tip width, and set a pressure, then issued a draw call with coordinates; it could smudge to blend colors the way you would rub a pencil with your finger, erase, or clear the whole canvas and start over. There were no solid fills — every area of shading had to be built up by layering strokes, exactly like real colored pencil. Then the same instruction to every model: reproduce the target as closely as you can, reviewing your own work as you go.
Crucially, the models were not flying blind. They could call view_canvas at any time to see their current drawing and compare it against the target, which is what makes this a test of self-correction and not just one-shot placement. Every tool call, every cent of API cost, and a similarity score for each finished piece were logged, so the experiment produced a paper trail, not just a gallery. That instrumentation is what turns "look, an AI drew a face" into something you can actually reason about — which model spent its effort where, what it cost, and whether all that reviewing made the drawing better or worse.
The drawings were scored with SSIM — the Structural Similarity Index, a 0-to-1 measure of how structurally close two images are, computed here after resizing both the drawing and the target to 256×256 pixels. On the Mona Lisa, the final SSIM scores were: Gemini 3.6 Flash 0.337, GPT-5.6 Sol 0.325, Claude Fable 5 0.286, and Grok 4.5 far back at 0.151. Read narrowly, Gemini won that specific image. But the authors were careful about what SSIM does and does not capture: it measures structural closeness, not how good a drawing looks to a human eye. A drawing can nudge the metric up by getting the broad light-and-dark blocking right while still looking nothing like a person.
On the human judgment that SSIM cannot see, the ranking flipped. The authors gave GPT-5.6 Sol the top spot for output quality across the whole batch — its Starry Night and its rose were among their favorites of the batch — despite Sol using fewer steps and fewer tokens than the models it beat. Gemini 3.6 Flash was solid but a step behind Sol. Claude Fable 5 landed in the middle. And Grok 4.5 was blunt-force worst: rarely producing anything usable, described in the writeup as "basically garbage." So the honest summary is two-layered — Gemini edged the pixel metric on the Mona Lisa, but Sol was the model you would actually want holding the pencil. That gap between a similarity number and a satisfying result is a theme worth carrying, because it shows up everywhere you try to automate visual quality.
Here is the finding that should stop anyone who assumes "spend more, get better." Totaled across the full set of drawings, Claude Fable 5 cost $160.58 — roughly twenty times the cheapest model — while Gemini 3.6 Flash came in at $12.87, Grok 4.5 at $9.21, and GPT-5.6 Sol at just $7.74. The most expensive run by a wide margin did not produce the best work, and the cheapest model produced the best human-judged work. Price did not track quality at all.
The gap was not raw per-token pricing so much as strategy. Claude leaned into heavy smudging and near-constant self-review, which meant a very long chain of tool calls and a large amount of vision processing on every review — each look at the canvas is another image fed back through the model. Sol did the opposite: it was terse and efficient, embedding its color and brush choices inline with the draw calls instead of spending separate turns on setup, and it finished in far fewer steps. The lesson for anyone budgeting real AI work is direct — the way a model uses tools can swing cost by an order of magnitude independent of the sticker price, and more effort is not the same as more quality. This is the same "pick the right tool, do not brute-force it" logic that governs a sane image and video workflow.
The most fascinating result is that the four models, handed identical tools, improvised four distinct working styles — which tells you their agentic decision-making differs as much as their output. GPT-5.6 Sol never bothered calling the separate set-color, set-brush, and set-pressure functions; it folded those parameters straight into its draw calls and moved fast. Grok 4.5 went the opposite direction, spending about 65% of its 1,349 total tool calls just setting parameters, averaging roughly 99 steps per drawing and still producing the worst results — a lot of fiddling, little to show. Claude Fable 5 was the blender: it emphasized smudging (around 123 smudge calls) and reviewed its canvas constantly. Gemini 3.6 Flash was the most self-conscious of all — nearly a third of its calls were view_canvas checks, looking at its own work more often than any other model.
This matters beyond curiosity. When four models diverge this hard from the same starting point, it means "which model is best" is genuinely task-dependent, not a single leaderboard — the terse-and-efficient one that wins a drawing might not be the one that wins a research task or a long-horizon coding job. It is the agentic-behavior version of the point made in how agentic AI works: give models tools and autonomy and their differences show up in how they act, not just in what they output. For anyone wiring these models into a pipeline, the takeaway is to test the actual behavior on your actual task rather than trusting a headline capability claim.
Everything above is interesting; this is the part that actually changes how you should use these models. The experiment tracked quality over the course of each run, and two things came out clearly. First, the benefit of self-review plateaued fast — similarity was essentially flat after about the fifth time a model checked its own canvas. Looking more did not help past that point. Second, and worse: every single scored run ended with a lower score than its own mid-run peak. The models reached their best frame partway through, and then kept editing, and made it worse.
Sit with that, because it is the load-bearing insight. These models can critique their own work well enough to improve early — they see a gap between canvas and target and close some of it. What they cannot reliably do is recognize the moment a piece is finished and stop. Left to run the loop themselves, they over-work it, adding strokes and smudges past the point of diminishing returns until the drawing degrades. The missing capability is not perception and it is not tool use; it is taste, or judgment — the "this is done, hands off" call that a human makes instinctively. That is not a quirk of drawing. It is the same reason unattended AI text keeps padding, unattended edits drift off-brand, and volume-first automation quietly rots, a pattern documented across why generic AI content stopped working. Capability without a stopping rule is not a finished product.
It is tempting to read a story like this as a horse race — Gemini won this, Sol won that, Grok flopped — and stop there. The more useful reading is about the shape of the gap. Generating an image is mature; executing a specific, controlled visual result agentically is not. The models disagree on strategy, cost varies by 20x with no quality guarantee, and none of them can be trusted to know when to stop. So the job of any serious content system is not to find the one god-model and let it freewheel. It is to route each task to the model that is actually good at it, and to keep a human decision at the point where judgment is still missing.
That is a very different posture from "prompt a model and post whatever it makes." It means treating raw model output as a draft that peaked somewhere and may already be past it, not as a finished asset. It means matching the format to the engine — a photographic scene, a face-locked avatar, a pixel-exact brand graphic, and a talking-head video are four different jobs that four different models do best, not one prompt sprayed at one model. And it means a person signs off before anything ships, precisely because the "when to stop" instinct is the thing the models demonstrably lack. Which is exactly the workflow the next section describes.
Kompozy is a content generation and multi-platform publishing engine — 18 output formats across video, image, and text, fanned to eight social platforms plus blog and email. Its whole design is a direct answer to the two failures this experiment exposed. First failure: no single model is best at everything, and the cheap one can beat the expensive one. Kompozy does not ask one model to freehand your brand. It routes each format to the engine that actually wins that job — GPT-image for photographic Photo Posts and infographics, Google Gemini's face-lock for consistent Persona Photos of the same avatar, HeyGen for talking-head video, and brand-exact HyperFrames for carousels and graphics that have to land pixel-perfect. HyperFrames is the sharpest contrast with the arena: instead of asking a model to draw your logo and layout stroke by stroke and hoping it does not drift, it renders the graphic deterministically from a template, so a month of posts reads as one brand instead of a scatter of off-model attempts.
Second failure, the important one: the models cannot tell when to stop, so they degrade past their peak. Kompozy builds the missing judgment back in as structure. A Persona Brief governs voice, positioning, and a banned-word list on every generation, so output is constrained toward your register rather than wandering. And Autopilot runs the cadence, but every piece passes a per-post review gate where a human approves it before it publishes — the explicit "this is done, ship it or send it back" decision the arena proved these models do not make on their own. That is not friction bolted on; it is the exact capability the experiment showed is missing, added at the one point where it matters. The result is throughput without the over-worked, off-brand drift that unattended generation produces.
Keep the framing honest, because overpromising is its own tell. Kompozy does not have better taste than you; it does not draw a better Mona Lisa. What it does is remove the two traps this experiment put a number on — using the wrong model for a job because it is the one you happened to prompt, and letting a model run its own loop past the point where the output was actually good. It picks the engine that wins each format and it keeps a person on the stop button. Generation, you already have from these models. The judgment about which model and when to stop is the part worth building a system around — and it is the part the pencil test showed they still cannot do alone. For the fuller picture of how these frontier models stack up as creative tools, see the Fable 5 vs GPT-5.6 comparison and the 2026 video AI model landscape.
Hand four frontier models a pencil and you learn more than any leaderboard tells you. They can all generate a Mona Lisa instantly, but drawing one exposed the seams: Gemini won the structural metric, Sol won the human eye and did it cheapest, Claude spent 20x more for worse work, and Grok floundered. They improvised four different strategies from one toolset, proving their agentic behavior differs as much as their quality. And every one of them peaked mid-run and then made its drawing worse, because none could tell when to stop. The capability is real and the judgment is not — which is precisely why finished, on-brand, shippable content still comes from matching the right model to each job and keeping a human at the point of "done."
In July 2026, TryAI ran a "canvas arena" that gave four frontier vision models — GPT-5.6 Sol, Claude Fable 5, Grok 4.5, and Gemini 3.6 Flash — a blank white canvas and a set of colored-pencil tools, then asked each to reproduce famous images including the Mona Lisa and Van Gogh's Starry Night stroke by stroke. The models could set a color, brush tip, and pressure, draw, smudge, erase, view the target, and view their own canvas to self-correct. Every tool call, cost, and similarity score was tracked. The point was not to make pretty pictures — it was to separate generating an image from executing one, which are different skills.
It depends on how you score it. By SSIM — a structural-similarity metric from 0 to 1 — Gemini 3.6 Flash edged the field on the Mona Lisa at 0.337, just ahead of GPT-5.6 Sol at 0.325, then Claude Fable 5 at 0.286 and Grok 4.5 far back at 0.151. But SSIM measures structural closeness, not how good a drawing looks to a person, and on human-judged quality the authors gave GPT-5.6 Sol the top spot overall, counting its Starry Night and its rose among their favorites of the batch. Grok 4.5 was described as producing "basically garbage." So Gemini won the pixel metric on that one image; Sol won the eye test across the experiment.
Wildly different amounts for similar or worse results. Across the full set of drawings, Claude Fable 5 ran up $160.58 — roughly 20 times more than the cheapest — while Gemini 3.6 Flash cost $12.87, Grok 4.5 $9.21, and GPT-5.6 Sol just $7.74. The most expensive model did not produce the best output, and the cheapest (Sol) produced the best human-judged work. The cost gap came from strategy, not price-per-token alone: Claude leaned on heavy smudging and near-constant self-review, piling up tool calls, while Sol embedded its color and brush settings inline with each stroke and finished in far fewer steps.
Only up to a point, and then they made things worse. Both drawing quality and the payoff from reviewing went essentially flat after about the fifth time a model checked its own canvas. More strikingly, every scored run ended with a lower similarity score than its peak mid-run — the models kept editing past their best frame and degraded it. That is the sharpest finding in the experiment: these models can critique their own work well enough to improve early, but they have no reliable sense of when a piece is finished, so left unattended they over-work it.
That generating an image and executing one stroke by stroke are different capabilities, and the second is still uneven. Given one identical toolset, the four models improvised four very different strategies — Sol was terse and efficient, Grok spent 65% of its calls just setting parameters, Claude obsessively smudged and reviewed, Gemini checked its canvas most often. That divergence shows the models have genuinely different agentic decision-making, not just different output quality. And the "peaks then degrades" pattern shows the missing piece is judgment — knowing when to stop — which is exactly the human-in-the-loop role that any serious production workflow still keeps.
No — it means drawing-by-hand is a deliberately hard test that is different from what these models are normally used for. Text-to-image generation is fast, cheap, and often excellent; the arena stripped that away on purpose to probe visual-spatial reasoning and agentic tool use, which are harder and less mature. The practical takeaway is not "AI can't make images." It is that raw model capability does not equal finished, consistent, on-brand output. Picking the right model for each job, and keeping a human gate that decides when something is done, is what turns a capable model into shippable work.
In July 2026, TryAI's "canvas arena" gave GPT-5.6 Sol, Claude Fable 5, Grok 4.5, and Gemini 3.6 Flash a blank canvas and colored-pencil tools and asked each to draw the Mona Lisa stroke by stroke. By SSIM, Gemini led at 0.337, then Sol (0.325), Claude (0.286), and Grok (0.151); on human-judged quality Sol won and Grok produced "garbage." Claude cost 20x more than Sol for worse results, and every run peaked mid-way then degraded — the models had no sense of when to stop. The lesson: generating an image and executing one are different skills, and judgment is the missing piece.
Get started → · ← All guides · Compare Kompozy vs other tools