Claude Opus 5 hits 30.2% on ARC-AGI-3, far ahead of GPT-5.6 Sol
The Decoder reports Anthropic's Claude Opus 5 scored 30.2% on ARC-AGI-3, well above GPT-5.6 Sol's prior 7.8% mark, and ahead of Fable 5. Developers say the model independently formulated reflection equations, a behavior not previously seen on that benchmark.
Key takeaway
The story that matters is not another Opus 5 launch recap but a rare step-change on ARC-AGI-3, where Anthropic's model more than triples the prior GPT-5.6 Sol record and shows novel self-derived reflection behavior.
Context
ARC-AGI-3 is positioned as a harder test of abstract reasoning than leaderboard chat benchmarks, so a jump from single-digit to roughly 30% is a structural signal for agent and reasoning stacks. The Decoder coverage frames Opus 5 as clearing both Anthropic's own Fable 5 tier and OpenAI's GPT-5.6 Sol on that specific measure.
For builders, the reported independent formulation of reflection equations matters as much as the headline score: it suggests less reliance on hand-crafted chain-of-thought scaffolding for some hard problems. That still leaves open how the result transfers to production cyber, tool-use, and long-horizon agent workloads outside ARC-AGI-3.
Numbers to know
- 30.2%Claude Opus 5 score on ARC-AGI-3
- 7.8%Prior GPT-5.6 Sol record on ARC-AGI-3 cited for comparison