Atlas

Benchmarks

← All benchmarks

BabaIsBench Stage 0: The Intro

Games · 2026-07-16

Quesma's BabaIsBench ports the puzzle game Baba Is You to the Harbor framework: an agent reads the board as JSON, plays through a simulator command line that runs the game's own engine, and submits a move string that the verifier replays through that engine; the instructions never name the game or the level. Stage 0, The Intro, is the game's eight opening levels. Every model gets three attempts per level in the Terminus 2 harness via OpenRouter, with a 60-minute limit per attempt, and the score is the share of its 24 attempts that are solved (pass@1). The stage is near ceiling: eight of the 32 models on the September 5, 2026 board solve every attempt.

Top models (higher is better)

ModelScore
Claude Fable 5.1100.0
Claude Fable 5100.0
Claude Opus 5100.0
Gemini 3.7 Flash100.0
Gemini 3.8 Flash100.0
GPT-5.5100.0
GPT-5.6 Sol100.0
GPT-6 Astra100.0
Opus 4.896.0
GLM-5.296.0
GLM-5.396.0
Grok 4.696.0
Kimi K396.0
Gemini 3.1 Pro Preview92.0
Gemini 3.5 Flash88.0
Gemini 3.6 Flash88.0
GPT-5.6 Terra88.0
DeepSeek-V4-Flash-073175.0
DeepSeek V4 Pro 081375.0
Grok 4.575.0
Sonnet 567.0
GLM-5.3 Flash67.0
GPT-5.6 Luna63.0
Qwen3.8-Max58.0
Hy354.0
MiniMax M346.0
Muse Spark 1.346.0
DeepSeek-V4-Pro29.0
Qwen3.6 27B29.0
Qwen3.7-Max (source unspecified)25.0
Qwen3.8 27B21.0
Loading Atlas data…