Cognition released FrontierCode 2.0 as a new coding benchmark. It draws tasks from 55% more repositories. Tasks now include security patches and performance optimizations. Grading criteria impose stricter correctness and design requirements. Reference solutions modify 67% more files and 89% more lines of code. Codebases are 25% larger on average. Thoroughness criteria grew threefold. The top model, Opus 5.5, scores only 45.5%. Scope failure rate dropped to 3.8%. The benchmark separates models more clearly across capability tiers.
What changed
FrontierCode 1.0 used fewer repositories, smaller codebases, and less stringent grading criteria with fewer blocking checks.
What it unlocks
Evaluating coding agents on realistic, mergeable tasks with stricter correctness, design, and thoroughness standards.
- 55% more repositories
- 43% more blocking criteria
- 67% more files modified on average
- 89% more lines of code modified on average
Sources