FrontierCode benchmark sets a new AI coding evaluation bar, with top maintainer approval at 13.4%
Cognition released FrontierCode, a coding benchmark built from 150 tasks by more than 20 open-source maintainers and judged against over 3,000 rules, with Claude Opus 4.8 reaching 13.4% approval in the hardest tier and GPT-5.5 reaching 6.3%.
Why it matters: Cognition's FrontierCode benchmark uses real maintainers and merge-to-main as the pass bar, and the best model clears only 13.4%.