Quesma tested a fresh group of language models on Baba Is Bench in August 2026. The benchmark asks coding-capable agents to solve levels of the puzzle game Baba Is You, where changing the rules on the board is part of solving the problem.
On the eight-level introductory stage, Gemini 3.7 Flash completed every trial in Quesma's pass-at-one measure. Grok 4.6 reached 96%, up from 75% for Grok 4.5. DeepSeek V4 Pro 0813 solved seven of eight levels at its first attempt, compared with three for the earlier V4 Pro in the same report.
The study also measured the cost of running those trials. Quesma reports an introductory-stage cost of about $5.38 for Gemini 3.7 Flash against $124.31 for Gemini 3.6 Flash. Its measured costs depend on the model, provider and agent procedure used for the experiment.
Not every newer model improved. GLM-5.3 matched GLM-5.2 on the introductory pass rate, while Qwen3.8-27B scored lower than Qwen3.6-27B in this particular game. The author cautioned that the Qwen difference was within experimental noise.
Quesma advanced models that cleared enough introductory levels to a harder stage called the Lake. Those results used different agent harnesses for some models, making a simple cross-model ranking less reliable. The article reports how the models handled this particular interactive puzzle, not their overall intelligence or performance on a real software project.