BabaIsBench Stage 1: The Lake
Games · 2026-07-16
Stage 1 of Quesma's BabaIsBench: the 15 levels of Baba Is You's second world, The Lake, including its two bonus levels, in the same Harbor port and engine-replay verifier as Stage 0. Only models that solve at least seven of the eight Stage 0 levels are run, once per level, each in a harness Quesma chose for it (Claude Code, Codex, kimi-cli or Terminus 2) with a 4-hour agent cap and no time limit in the instructions. The score is the share of the 15 levels solved in that single run; runs that Quesma stopped by hand because of their cost count as unsolved.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5.1 | 100.0 |
| GPT-6 Astra | 100.0 |
| Claude Fable 5 | 93.3 |
| Claude Opus 5 | 93.3 |
| Gemini 3.7 Flash | 93.3 |
| GPT-5.6 Sol | 93.3 |
| DeepSeek V4 Pro 0813 | 86.7 |
| Gemini 3.8 Flash | 86.7 |
| GPT-5.5 | 86.7 |
| Grok 4.6 | 86.7 |
| Opus 4.8 | 80.0 |
| GLM-5.3 | 80.0 |
| GLM-5.2 | 73.3 |
| Kimi K3 | 73.3 |
| GLM-5.3 Flash | 60.0 |
| Gemini 3.1 Pro Preview | 53.3 |
| Gemini 3.5 Flash | 46.7 |
| DeepSeek-V4-Flash-0731 | 40.0 |
| Gemini 3.6 Flash | 40.0 |