Sokoban eval for LLMs: gpt-6.2 sol lands #2, astra #3; humans invited to beat model routes

banteg · x · 2026-11-18

Developer banteg launched a Sokoban puzzle benchmark where language models play Jonah Ostroff's 66-room trilogy, submitting moves or full-room plans each turn, replayed in the original PuzzleScript game. gpt-6.2 sol and astra (medium effort) both cleared it, ranking #2 and #3. The site also challenges humans: models hold shortest-route records in every room, but shorter routes exist.

Related event: banteg's Sokoban Benchmark: GPT-6.2 Models Clear All Levels(3 posts)→

Original post →

More from Models

Models channel →