Playground · Animation · Canvas · Standing test: we add each new model as it launches
Three AI models paint a lemon on a blue plate in thick oils, stroke by stroke, in one HTML file. Watch every answer live and rake the light.
Thick oil paint is about texture as much as the picture: brush marks, ridges of impasto that catch a raking light, and an order to the painting, background first and highlights last.
Each answer has to make the piece from start to finish in one HTML file, hold on the result, and let the pointer move the light, an idea we took from @kaolti's test of AI models embroidering a logo. Claude Opus 5.5 and Muse Wonder each had one try over the API at their default settings, with room for 128,000 tokens. Muse Wonder Full is not on the API yet, so it answered the same prompt in Muse Code, Meta's agent app, where it could screenshot its own page and fix what was wrong before handing it over. That is a different test from a single try, closer to how people use Full today; it has no metered time or cost, and its note says how many rounds it took.
Paint a lemon on a blue plate as a thick oil painting on canvas, stroke by stroke, with visible brush marks and impasto. Make it a single self-contained HTML file (inline CSS and JavaScript, no external libraries, fonts or images) that fills the window. Show the piece being made from start to finish, then hold on the finished result. Moving the pointer moves the light, so the surface catches it the way the real material would. Click to start again.Show the whole prompt
Default settings: no thinking effort set, no temperature, and as much room as each model has (its own output limit). A card says so when an answer was made with other settings: “limit 128k” means it could use up to 128,000.
| Rank | Model | Checks | Time | Cost |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 Our pick | 4 of 4 | 7.2 min | $0.897 |
| 2 | Muse Wonder Full | 3 of 4 | – | – |
| 3 | Muse Wonder | 3 of 4 | 2.6 min3 tries | $0.300 |
Ranked by our pick, then checks passed, then cost. Time and cost count every try; each dot is one check, in the order listed below (filled: passed, half: partly, empty: failed). Click a model for its notes.
Paints the whole canvas stroke by stroke: a warm ground, a grey-green background in crossing strokes, brown table strokes, a plate in swirls of blue, then the lemon in layered yellow impasto. The ridges catch the light and shift as you move the pointer. The background is busy, but this is the one that looks like oil paint. 7.2 minutes, $0.90.
Made in Muse Code, not over the API: Full could screenshot its own work and fix it, and took 3 rounds. A framed canvas painted in a sensible order: the background blocked in with short horizontal dashes, the plate laid in, the lemon stroke by stroke, then the leaf and highlights last. But the lemon is small and breaks up into scattered yellow dashes on a nest of blue strokes, and the ridges barely change as the light moves.
A dark streaked background, a wooden table, a blue plate with white rings and a glossy lemon with a long green leaf, mostly in place within eight seconds. The pointer drags a large glare across the plate. It looks like a rendered 3D lemon with a paint texture laid over it rather than an oil painting. Our first two requests were refused before it wrote anything (by our account's credit limit at its listed 943,718 tokens of room, then over a thinking setting this endpoint does not take); the third, with 128,000 tokens, is the answer. 2.6 minutes, $0.30.
One answer per model, at each model's default settings unless the card says otherwise, through OpenRouter. If an answer ran out of room or stopped short, we asked again with more room, then with thinking off, and say so. This is a quick look, not a benchmark: run it again and every answer will differ. For measured results, see our benchmarks.
How we pick. Claude Opus 5.5 tried every answer (clicking, dragging and playing it in a browser), marked the checks and drafted the notes. Our pick is a judgement among the answers with the most checks passed, not a score: the checks show what works, the pick is which of those we would rather use. Opus is an Anthropic model, and Claude answers are our pick in 17 of 28 tests, so weigh Claude picks with that in mind. GPT-6.1 Sol, shown anonymised stills and our check marks, put Claude Opus 5.5 first for brushwork that reads as thick oil paint, then Muse Wonder, whose texture it found glossy and repetitive, then Muse Wonder Full. We agree.
What we checked, by trying each answer ourselves: Paints in a real order: background and plate first, then the lemon, stroke by stroke; Thick paint and brush marks show, and the ridges catch the light as the pointer moves; The finished picture clearly reads as a lemon on a blue plate; Finishes, holds on the finished piece, and starts again on a click. A partial pass counts as half.
Claude Opus 5.5's lemon is the one with real impasto: layered strokes with ridges that catch the light as you move the pointer, on a plate painted in swirls of blue. Its background is busy, which keeps this close. Muse Wonder's picture is cleaner but looks rendered rather than painted, and Muse Wonder Full, after three rounds in Muse Code, paints in the right order but ends on a small, scrambled lemon.
Our pick: Claude Opus 5.5
Our pick: Claude Opus 5.5
Our pick: Claude Opus 5.5