Cognition introduces FrontierCode 2.0

Engineering judgment beyond a specification

Cognition introduced FrontierCode 2.0 on October 27. Its tasks include security patches and performance improvements, with deliberately underspecified requests intended to test engineering judgment as well as functional correctness. Open-source maintainers helped construct the tasks and grading rubrics. [1]

Cognition reports broader repository coverage, larger changes and more demanding grading than its earlier benchmark. The results measure performance under Cognition's evaluation rules; Artificial Analysis identifies the test set as private, so the release does not provide an openly downloadable task corpus. [1][3]