Claude Sonnet 5.5 scores 72.4% on ARC-AGI-3 with ARC Prize's Provider Adapter harness
ARC Prize's verified testing puts Claude Sonnet 5.5 at 72.4% on ARC-AGI-3, a benchmark of 25 interactive game environments. That score used the Provider Adapter harness, which keeps the model's reasoning state between requests and compacts long conversations so it can reuse prior work. With the Standard harness, which only carries forward notes the model chooses to keep, it scored 24.3%. Sonnet 5.5 also reached 83.9% on ARC-AGI-2 and 96.5% on ARC-AGI-1.