The next useful AI coding benchmark will not ask whether the model can write code. It will ask whether a maintainer would merge it.
Cognition's FrontierCode is built around that question, and it is a better direction than another leaderboard on generated snippets.
In real teams, "the code works locally" was never the finish line. A change still has to fit the codebase, respect its constraints, survive review, avoid hidden maintenance debt, and stay legible to the people who will own it later. FrontierCode tries to measure exactly that. The tasks were built by 20+ open-source maintainers on the repos they actually own, defining what "mergeable" means in their codebase, graded on correctness, test quality, scope discipline, and style. Top models still score in the low teens on the hardest set.
The old benchmark asked "can it solve the task?" The better one asks "should we accept the change?"
Writing code was the part we already knew how to automate. Owning the consequence of the change is the part that still needs a human in the loop.