Cognition says GPT-6 Astra reached near-Fable 5 coding quality (within 0.4 points) at 64% lower rollout cost on FrontierCode 1.1

Now, FrontierCode benchmark is quite unusual because it asks whether an AI coding agent can produce a pull request a maintainer would actually merge, rather than stopping at functional correctness.

Cognition built 150 tasks with maintainers from 36 open-source repositories, grading correctness, regression safety, tests, scope, style, and adherence to each codebase's conventions.

The score is a weighted rubric aggregate rather than a solve rate, and any run that misses a blocking requirement receives zero.

METR separately found that roughly half of earlier SWE-bench Verified patches that passed automated tests still would not have been merged by maintainers.

i.e. FrontierCode really measures review-quality code.