Tools
An agent skill claims Claude Code on Opus 5 cleared all 25 public ARC-AGI-3 games at 100.00 RHAE in 7,645 actions
pbshgthm/arc-skill, created 2026-08-19 and at 78 stars, is an agent skill for playing ARC-AGI-3 built on one rule: the agent must state what an action will do before spending it. The repo claims Claude Code on Opus 5 finished all 25 public games at 100.00 RHAE in 7,645 actions. The result is self-reported in the repo description with no independent confirmation found, and the repo has not been pushed since 2026-08-19, so treat the number as a claim rather than a verified benchmark. The predict-before-acting constraint is reusable regardless of whether the score holds up.
Source
↳ Follow the thread