10%20%30%40%
7.8%Record listingARC Prize · semi-private set13.3%Official harnessOpenAI’s run · public set38.3%+ preserved reasoning+ context compaction · public set
3× the score, same weights, same set
GPT-5.6 SOL ON ARC-AGI-3
Same model throughout. The record was set on a different task set; the last two share one and differ only in the harness.