New benchmark tests LLM reasoning on Baba Is You puzzle levels

Researchers evaluated Claude Fable 5 and GPT-5.6 Sol on the Baba Is Harbor benchmark, finding that while models can solve most levels, they are often less cost-effective than humans. Experiments revealed significant cost variations, with Gemini 3.5 Flash costing 2.4x more than Fable 5 to solve the game's introductory stage.