- Sources: ARC Prize results, HN discussion
- Summary: ARC Prize published verified results for OpenAI GPT-6 Astra on 2026-09-02. The best ARC-AGI-3 Semi-Private score is 99.95%, at high reasoning on the Provider Adapter harness, at a cost of $18,817. The same model scores 62.71% at max reasoning on the Standard harness, at a cost of $26,098.
- Why it matters: The same model and benchmark differ by more than 37 points across two harnesses, so a reported ARC-AGI-3 score means nothing without the harness named alongside it.
- Follow-up: What ARC Prize documents about the Provider Adapter harness surface, and whether other verified models show the same harness spread.
send feedback on this story