Definition
Anthropic-promoted software-engineering / agent evaluation suite used to claim state-of-the-art results for Claude Opus 5 (July 2026).
Key Points
- Distinct from independent arc-agi-3 adaptation results verified by arc-prize for claude-opus-5 (2026-07-26-arc-prize-claude-opus-5-results)
- Cited as SOTA for Claude Opus 5 vs prior Opus/Fable on coding agent tasks (2026-07-24-anthropic-claude-opus-5-launch)
- Company-run numbers; independent third-party replication not required for marketing claims
- Related coding/agent benches: CursorBench, AutomationBench, OSWorld