Introducing FrontierCode

Cognition has released FrontierCode, arguing that existing coding benchmarks are no longer adequate for measuring the real production capabilities of AI coding agents. The article notes that previous evaluations like SWE-Bench Verified and SWE-Bench Pro primarily tested functional correctness — but once models can write code that passes tests, the more important question becomes: would this PR actually be merged by a maintainer? FrontierCode therefore shifts the evaluation focus to mergeability, which encompasses end-to-end code quality including behavioral correctness, regression safety, mechanical checks (build/lint/style), whether agent-written tests are effective, whether the scope of changes is appropriately restrained, and whether the code conforms to project conventions and is readable.