FrontierCode v1.1 (Main)

Agentic coding

Cognition’s FrontierCode evaluation tests whether models can complete difficult coding tasks while meeting the standards of high-quality production codebases — code that is correct, maintainable and reviewable, not merely passing tests. The v1.1 Main set is 100 tasks derived from real pull requests (Cognition reports a separate 150-task Extended set), graded on held-out unit tests plus weighted code-quality rubric criteria, so an out-of-scope change can cost a model points even when it works.