FrontierCode 2.0
Cognition's second-generation code-mergeability benchmark (released 2026-10-27): 132 private tasks crafted by the open-source maintainers of the repositories they come from, graded against maintainer-defined rubrics where any failed blocker criterion scores the solution 0. Compared with FrontierCode 1.1, tasks come from 55% more repositories and add categories such as security patches and performance optimizations; reference solutions touch 67% more files and 89% more lines in codebases 25% larger; and tasks carry 43% more blocking criteria, with a new emphasis on 'thoroughness' (fixing every instance of a bug, matching existing support for related inputs). In 80% of tasks the reference solution was corrected from the upstream PR. Each model is run 5× at every available reasoning effort; runs that consulted solution-bearing sources are zeroed.
What this benchmark measures
Cognition's second-generation code-mergeability benchmark (released 2026-10-27): 132 private tasks crafted by the open-source maintainers of the repositories they come from, graded against maintainer-defined rubrics where any failed blocker criterion scores the solution 0. Compared with FrontierCode 1.1, tasks come from 55% more repositories and add categories such as security patches and performance optimizations; reference solutions touch 67% more files and 89% more lines in codebases 25% larger; and tasks carry 43% more blocking criteria, with a new emphasis on 'thoroughness' (fixing every instance of a bug, matching existing support for related inputs). In 80% of tasks the reference solution was corrected from the upstream PR. Each model is run 5× at every available reasoning effort; runs that consulted solution-bearing sources are zeroed.
Rows on this page are sourced from public benchmark artifacts, leaderboard exports, or source-linked model reports. Each row keeps benchmark version, source model name, and available run details attached to the score.
The metric shown here is weighted score (Main). It should be interpreted within FrontierCode 2.0, not compared as part of a site-wide ranking.
What to be careful about
A rebuilt task suite, so scores are not comparable with FrontierCode 1.1 (the frontiercode column). Opus 5.5 scores 54.6 on 1.1 and 47.3 on 2.0, and SWE-2 scores 50.0 and 24.1. Each model is shown at its best-scoring effort with the harness Cognition ran it in, which is named in each row.
Frequently asked
What is FrontierCode 2.0?
Cognition's second-generation code-mergeability benchmark (released 2026-10-27): 132 private tasks crafted by the open-source maintainers of the repositories they come from, graded against maintainer-defined rubrics where any failed blocker criterion scores the solution 0. Compared with FrontierCode 1.1, tasks come from 55% more repositories and add categories such as security patches and performance optimizations; reference solutions touch 67% more files and 89% more lines in codebases 25% larger; and tasks carry 43% more blocking criteria, with a new emphasis on 'thoroughness' (fixing every instance of a bug, matching existing support for related inputs). In 80% of tasks the reference solution was corrected from the upstream PR. Each model is run 5× at every available reasoning effort; runs that consulted solution-bearing sources are zeroed. It is a coding benchmark measured by weighted score (Main).
What does weighted score (Main) mean on FrontierCode 2.0?
FrontierCode 2.0 reports weighted score (Main) (%); higher is better. Scores are shown only within FrontierCode 2.0 and are never averaged with other benchmarks.
What is the top reported FrontierCode 2.0 score?
Claude Fable 5.2 has the top reported score on FrontierCode 2.0: 50.2% (weighted score (Main)).
Why do FrontierCode 2.0 scores differ across runs?
Harness, scaffold, reasoning effort, and prompt setup change results, so two runs of the same model can differ. evals.report keeps each score with its run context so the differences stay visible.
Does evals.report rank models across benchmarks?
No. FrontierCode 2.0 scores are shown within their own metric; evals.report never combines benchmarks into a composite ranking or a single "best model".