Atlas

Benchmarks

← All benchmarks

Gym-Anything-RT (non-real-time counterparts)

Games · 2026-10-27

The 100 non-real-time counterparts of the Gym-Anything-RT minigames: each poses the same task, controls, information and difficulty as its real-time minigame, minimally changed in timing so that acting in real time is no longer required. Run live at difficulty 5 with full interaction under the same screenshot, mouse-and-keyboard and step limits, one attempt per task; the score is the share solved, in percent.

Top models (higher is better)

ModelScore
GPT-6 Astra66.0
Claude Opus 5.557.0
GPT-6 Luna17.0
Qwen3.8 27B2.0
Qwen3.5-9B1.0
Loading Atlas data…