Claude Opus 5 solves an ARC-AGI-3 game by turning vision into linear algebra
daniel_mac8 · x · 2026-07-27
Claude Opus 5 reportedly solved one ARC-AGI-3 game by using vision to turn the puzzle into a linear-algebra problem.
- The attached ARC Prize Foundation figure shows Opus 5 (High) reaching 30.16% RHAE at high effort, far above the other models plotted on the cost/performance chart.
- The foundation’s judge review says the model completed Axis Reflect (ar25) with a perfect 100, clearing all eight levels in 294 actions.
- Its key strength was described as converting the visual puzzle into explicit algebra: by level 2 it had derived a reflection equation, and by level 8 it generalized to two-dimensional reflections, partitioned 60 targets into mirrored quadrants, and planned each piece’s destination before acting.
- The judge also noted some hypothesis churn and repetitive narration, but said failed exploration seemed to update understanding rather than create loops. The conclusion: once the right ontology is found, execution is extremely reliable.
Related event: Claude Opus 5 Sets New ARC-AGI-3 Record(15 posts)→
More from Models
- Is DeepSeek's rumored K3 a scaled-down model, or something bigger? X users debate — teortaxesTex · 2026-09-11
- DeepSeek update keeps cache hits mid-conversation, cuts costs 36.6% — teortaxesTex · 2026-09-11
- AI Sextet offers 6 models free and unlimited for 14 days, including DeepSeek and Qwen — airesearch12 · 2026-09-11
- BullshitBench update: GPT-6-Astra beats all prior OpenAI models but still trails Anthropic — scaling01 · 2026-09-11
- Astra Scores 83% on GauntletBench, First Computer-Use Agent to Beat Human Baseline — ducha_aiki · 2026-09-11
- Kimi K2.8 Preview rolls out: near-K3 coding performance, 1M context for all tiers — teortaxesTex · 2026-09-11