Researchers released Kepler, an open‑source tool that builds and checks executable world models for the ARC‑AGI‑3 benchmark. Running under a frozen Claude Opus 5 configuration, Kepler earned a server‑verified 100.00 RHAE on every one of the 25 public games. It did this without per‑game model selection or rerunning with score conditions, and on most levels it used no more actions than a median human player.
Why it matters
For AI developers, the result shows that a single world‑model approach can match human‑level efficiency across many tasks without costly per‑task tuning.
