← All benchmarks

MazeBench ASCII with Python

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 track the agent also gets a run-scoped Python scratchpad (shell, web and host files blocked) and may write solvers that simulate the game. These are separate runs from the no-Python headline. The score is the number of the world's 100 hidden gems collected; Atlas imports only settled runs on the current 100-gem world. 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)