Atlas

Benchmarks

← All benchmarks

MazeBench ASCII

Games · 2026-07-28

Maze Bench's open world of more than 200 rooms, ordered as a curriculum, filled with Sokoban-style box puzzles plus elevators, toggle walls and icy slopes, rendered as isometric ASCII (4x4 characters per tile) with 20 rotatable camera angles. An agent plays a single open-ended run in its vendor coding harness (Codex, Claude Code, Grok Build, Muse Code and others) through an MCP server with 11 actions: four moves, four camera rotations, undo, level reset and teleport to a visited room. There is no action limit; the operator stops runs on stagnation, usage limits or cost, so a score reflects one run of uncontrolled length. In this headline track Python is disabled, so the agent must plan in-game. The score is the number of the world's 100 hidden gems collected (one point per gem). Atlas imports only settled runs on the current 100-gem world; earlier July-August runs played smaller 71-90-gem revisions. It is a diagnostic, not a capability family: each model has one open-ended run that the operator stops on stagnation, usage limits or cost, many runs are scored while failed or paused, and gems are collected sequentially in one run rather than as independent items.

Top models (higher is better)

ModelScore
GPT-6 Astra16.0
GPT-6.1 Sol9.0
Claude Fable 5.12.0
Claude Haiku 5.51.0
Claude Sonnet 5.51.0
GPT-6 Sol1.0
Grok 4.71.0
GPT-6 Luna0.0
Muse Spark 1.20.0
Muse Spark 1.30.0
Muse Wonder0.0
Loading Atlas data…