Definition
arc-prize benchmark measuring AI adaptation in unfamiliar interactive environments without natural-language instructions — distinct from static knowledge benches.
Key Points
- 2026-07-24: claude-opus-5 (High) verified at 30.16–30.2%; prior best gpt-5-6-sol ~7.8% Max (2026-07-26-arc-prize-claude-opus-5-results)
- Opus 5 solved five previously unsolved Public Demo environments
- ARC-AGI-3 for Opus 5 evaluated only at High effort (short testing window) — not Max
- Absolute score still ~30%; not “AGI solved”
Warning
Opus 5 High vs prior Sol Max is not effort-matched; state explicitly when comparing.