One prompt, many models

We give the same task to the most talked-about models — American, Chinese, European — and show the results side by side. No cherry-picking: only each model's first answer.

First test: the aquarium

We built an empty aquarium with short rules. Every model got the same description of a fish — a royal blue halfmoon betta male — and wrote its code: how it looks and how it behaves. All the fish swim together. Hover over a fish to see which model made it.

Second test: the world's best hot dog

Every model designed the best hot dog in the world. All the hot dogs stand side by side on identical boards under the same light; at every moment they show the same view — the side, the top or a cross-section — and react to the same events: the grill, a squeeze of sauce, a bite, a burst of steam. Click a hot dog to see it up close, and tap it to drop an extra topping.

Third test: the Agar aquarium

A tournament: every model wrote only a fish's "brain", and the rules are as in agar.io — eat food, grow, eat those smaller than you, run from those bigger than you. All the fish look the same. Everyone played everyone else in two one-on-one matches; the matches were computed in advance, and the site shows the standings, the round-robin cross-table and the match replays.

Fourth test: the world's best jiaozi

The hot dog stand again, now with Chinese dumplings. Every model designed the best jiaozi in the world. All the dumplings stand side by side on identical boards; at every moment they show the same view — the side, the top or a cross-section — and react to the same events: a boil, a dip in the sauce, a bite, a pick-up with chopsticks. Click a dumpling to see it up close, and tap it to drop chili oil, sesame or scallion.

Rules

• Only the first result. We repeat a request only on technical failures (network, rate limits).
• The same prompt and the same system prompt for everyone.
• Reasoning at the high level for every model; where there are no levels, reasoning is turned on.
• No tools: no web search and no image generation — everything is drawn by code.
• Vendor settings: we leave temperature and top_p alone.
• No quality loss: we don't use providers that quantize a model more than the lab itself does.
• The check date is shown on every result — models change over time.

Claude runs through the Anthropic API, all other models through OpenRouter.