Six models play real-time snake: Jev hits 28 of 34 moves on time
In a snake game with 400 ms per move, Jev made 28 of 34 moves on time and ate 3 apples, while five chat models all crashed on move 6; Jev's cost of $19.24 per 1M moves was second only to Haiku.
Claude Opus 5.5 didn't make a single move. A model that can't even write a word ate 3 apples. I put 6 AI models in a real-time game of snake. 400 ms per move, 1 API call per move, same board, same seed. Miss the window and your snake keeps going straight. Moves on time: Jev: 28 of 34, ate 3, 331 ms per move Claude Haiku 5.5: 0 of 6, 1,423 ms GPT-6 Sol: 0 of 6, 1,560 ms Gemini 3.8 Flash: 0 of 6, 1,724 ms Claude Sonnet 5.5: 0 of 6, 2,364 ms Claude Opus 5.5: 0 of 6, 3,437 ms Five chat models hit the wall on move 6 without making one turn. Cost per 1M moves: Haiku $17.98 · Jev $19.24 · Gemini $94 · GPT-6 Sol $232 · Sonnet $353 · Opus $706 The smartest model in the race was also the slowest and the most expensive. Is your agent making decisions, or still thinking about them?