Model reports

FrontierCode

Cognition's production-standard coding evaluation, asking whether models write *good* code rather than merely correct code. Unrelated to Terminal-Bench's Frontier-Bench despite the shared prefix. Reported per reasoning-effort setting (the Diamond subset at xhigh), so the effort level has to travel with the number. First-party to a competing coding agent's vendor, and revised as 1.1 on 2026-07-07, so the methodology version has to travel with the number too.