[AINews] FrontierCode: Benchmarking for Code Quality over Slop

Cognition's FrontierCode is a new coding benchmark where every task required 40-plus hours of work from open-source maintainers. It asks a question SWE-bench never did: would you actually merge this code? Models have been gaming correctness metrics while producing unmaintainable slop. FrontierCode measures that gap directly.

The benchmark arrives after METR documented that many SWE-bench-passing PRs would not survive a real code review, and after SWE-bench Pro still failed to enforce quality rubrics. FrontierCode's design is explicitly modeled on FrontierMath, which held frontier models below 2% success on hard problems in late 2024. Its third and hardest tier maps the capability jump that emerged around December 2025, the same acceleration that made agentic coding loops viable at scale.

The full report is worth reading for the rubric construction alone. How Cognition operationalized 'mergeable code' into a scorable signal is the methodological question the whole field has been avoiding. That detail lives in the original, not here.