Frontiercode

FrontierCode is Cognition's test of coding agents, and it asks whether a project's maintainers would actually merge the agent's patch, not only whether the tests pass. Cognition launched it in June 2026 with 150 tasks from 36 open-source repositories, built by more than 20 of those projects' maintainers. Version 1.1 came out in July 2026 and is still the current version as of October 2026.

Each task carries criteria its maintainer wrote: tests, build and lint checks, limits on which files and how many lines a patch may touch, and an LLM review of the code. Some are blockers, the problems a reviewer would refuse a pull request over, and failing one scores the attempt zero. The leaderboard reports a score and a pass rate on Main, the 100 hardest tasks, and on Extended, all 150. The top score on Main as of October 2026 is about 55%.

The tasks are private, so you can't run it on your own agent. Cognition runs the official leaderboard and ranks its own model on it. Artificial Analysis runs a separate version with every model in one shared harness, and it ranks some models in a different order.