← All benchmarks

Gym-Anything-RT (real-time, live)

Games · 2026-10-27

Carnegie Mellon University's benchmark of 100 visual minigames that require real-time interaction, generated and audited by coding agents from a six-capability taxonomy (for example, clicking the keys of a code on a keyboard that rotates in 3D between observation and action). The environment keeps running while the model reasons and acts, and models see only screenshots and act only through mouse and keyboard input, with at most 100 steps per attempt and no code execution. Scores are the share of the 100 tasks solved in one attempt each at the highest of five difficulty levels with full (direct) interaction, in percent; task-specific verifiers grade each attempt.

Top models (higher is better)