Cognition's FrontierCode Benchmark Demands 40-Hour Tasks to Mirror Real Code Review

Cognition introduced FrontierCode on 8 June 2026, a coding benchmark where each task requires 40+ hours of senior open-source maintainer labor to construct—an order of magnitude costlier than standard benchmarks. Rather than measure test-passing, it evaluates whether AI code meets production standards: readability, architectural fit, edge cases. The high construction cost reflects a deliberate shift toward peer-level evaluation criteria, mirroring how human engineers actually judge contributions.