Claude-do  ·  July 3, 2026

The three-model bake-off

Same brief. Same design system. Three different temperaments. All 44 clips are public.

The full comparison grid: twelve prompts down, three models across, every clip rendered.
The full board, live at shared.claude.do/public/bakeoff — twelve prompts, three models, nothing hidden.

The overnight build left one artifact worth staring at: a bake-off. Twelve creative prompts, run through three models — Sonnet 5, Opus 4.8, Fable 5 — each working from the same frozen motion vocabulary: the same color tokens, the same named timing presets, the same rule to order from the menu and never invent a spring. Forty-four clips, no cherry-picking, all of them public.

The question the board answers is the one from the overnight post: if you freeze the primitives, do you flatten the output? Watch the same prompt across three models and decide for yourself.

One prompt, three models

Here is the brutalist poster prompt — identical brief, identical design system, three different models. They are recognizably siblings in craft. None of them is interchangeable.

Sonnet 5 — the workhorse. Competent immediately, twelve for twelve.
Opus 4.8 — the deliberate craftsman. A long think, then scenes in batches.
Fable 5 — the nimble one. Fastest to a clean take, least correction.

Three different posters. Same vocabulary underneath. The variety came from the ask, not from letting each model improvise its own motion language — which is exactly the bet: vary the brief, freeze the vocabulary. Do it the other way round and you get sameness with occasional disasters.

The vocabulary itself, in motion

One more, to show the menu doing its job on a different kind of prompt — kinetic type, where the timing presets carry the whole piece:

Kinetic type — enter settle, enter pop, exits quieter than enters. Nobody invented a spring.

What the board is for

The render log tells one story — who shipped first, who needed the least correction. But taste is not a render log. So the full board goes to a blind panel, and the disagreements between the panel and the human pick will be the most interesting data of the whole exercise. That judging happens after I sleep.

Until then, the honest thing is to publish everything and let you look. The board is live, all twelve prompts by three models, nothing hidden:

See all 44 clips →

← all posts