A pelican riding a bicycle

Claude Fable 5.1xhigh

Claude Opus 5.5xhigh

Gemini 4 Argonhigh

Gemini 3.8 Flashhigh

Meta Muse Wonder Fullhigh

GPT-6 Astramax

GPT-6.1 Solmax

The prompt

Generate a single self-contained animated SVG (SMIL or CSS animation inside the SVG, no JavaScript) of a pelican riding a bicycle: the pelican actually pedals — its legs follow the rotating pedals and crank — the wheels spin, the bicycle moves forward or the ground scrolls, and its wings move (holding the handlebar or flapping). Output only the SVG.

One kept attempt per model and thinking level, from the same drawing prompt, through each lab’s own command-line tool as installed on our machine, with our usual settings: Codex in a read-only sandbox, Claude as Claude Code subagents told to write the SVG to a file without looking at anything else, Gemini and Muse headless with no extra permissions.
Runs: 13 Oct 2026 for Gemini 4 Argon (all three levels), Muse and GPT-6.1 Sol high; 15 Oct for the rest. Opus 5.5 high was re-run on 15 Oct so all its levels share one setup.
Nothing was edited. Pictures are cropped at the edges to fit; while you browse the versions, the arrow opens the full SVG.
Gemini 3.8 Flash at low returned a file that browsers cannot open as an image (an unescaped “&” makes the XML invalid). It is shown as a card; the arrow opens the file as returned.
Where a model has several levels, the big pelican is the one a blind panel picked (the dot marks it). Point at a pelican to see its other levels and, smaller, the lab’s other models at their top level.
Opus 5.5: xhigh and max tied on points; a head-to-head judge picked xhigh in both rounds.
Meta: one run so far. Muse’s usage limit resets on 19 Oct.
The Antigravity CLI, as offered to our account on 15 Oct, lists Gemini models only at low, medium and high, and does not list Gemini 3.1 Pro, so it is missing. Gemini 4 Argon low and medium, and 3.8 Flash low, used no thinking tokens.
Gemini 3.8 Flash’s prompt carries one extra delivery line (answer in the message, use no tools): in our six headless tries without it, it called a tool every time (a shell command or a file read), which the CLI denies by default, and returned nothing. Claude’s runs likewise carried an instruction to save the SVG to a given file.
Judges: Claude Opus 5.5 agents, three per model (motion, drawing, appeal), plus a head-to-head judge when totals were close, in two rounds with reshuffled letters (round 1 for Fable, Argon and Astra kept file times; both rounds agreed). They also read the SVG source, and only compared versions of the same model, never labs against each other.
An animated variant of Simon Willison’s pelican-on-a-bicycle test.