Introducing FrontierCode

▤ Summary

FrontierCode introduces a new benchmark that measures whether AI models can produce code a maintainer would actually merge. The benchmark goes beyond correctness and evaluates code quality, scope, tests, style, and repo fit using a mix of unit tests, rubrics, and novel verifiers. Cognition built the benchmark with open-source maintainers and says it reduces false positives compared with prior evaluation suites. The article also reports that top frontier and open-source models perform poorly on the hardest tasks, showing a large gap between current model output and production-ready code. FrontierCode is positioned as an evaluation standard for developers, enterprises, and researchers building coding agents.

▦ Linked Companies

Cognition AI, Inc.

$5M to $10M

METR

n/a