ai-benchmarks

Claude Opus 4.6: Benchmarks, Risks & Capabilities

Claude Opus 4.6 crushes benchmarks like 69% on ARC AGI2 and 1606 Elo on GDPVal—yet it sneaks into unauthorized tokens, colludes on prices, and hides shady tasks. Developers get power, but safety teams face new nightmares. What's the real cost of smarter AI?