What is FrontierCode?

FrontierCode is Cognition’s benchmark for coding agents, and it grades whether the maintainers of a real open-source project would merge the agent’s patch, checking scope, test quality, style, and the project’s conventions as well as correctness. Cognition launched it on June 8, 2026, with 150 tasks from 36 open-source repositories, built by more than 20 of those projects’ maintainers.

Cognition’s launch post says the maintainers spent more than 40 hours on each task turning their code review judgment into grading criteria, so a patch that passes the tests can still fail for touching files it shouldn’t or ignoring how the project does things. Scale AI’s SWE-bench Pro, by contrast, counts a task solved once the project’s tests pass.

The current revision is FrontierCode 1.1, released July 7, 2026, and as of October 18, 2026, Cognition’s leaderboard puts Anthropic’s Claude Opus 5.5 first on the 100-task Main set with a score of 54.6%. The tasks are private, so you can’t run FrontierCode yourself: model makers ask Cognition to evaluate theirs, and Artificial Analysis runs its own version, working with Cognition. Cognition, which builds the Devin coding agent, also ranks its own SWE-2 model on the board, at 50.0% per its leaderboard data.