What changed in the latest version of FrontierCode?

Cognition’s FrontierCode 1.1, released July 7, 2026 and still the current revision as of October 18, 2026, zeroes any run that looks up a task’s upstream fix online, demotes 75 overly strict blocker criteria to non-blockers, and stops reporting the Diamond subset. Cognition’s leaderboard changelog lists only new models and reporting changes since then.

The internet rule is the main change. FrontierCode’s tasks come from real pull requests in open-source repositories, so the fix can exist publicly, in the upstream repository, a mirror, or the package’s latest release. Cognition kept internet access, since some tasks need it by design to look up API contracts, and added a prompt defining fair use plus a check that flags any run referencing the source pull request or upstream patch. Cognition’s 1.1 post says the prompt alone keeps unfair use below 1% for every model it evaluated, and the leaderboard has shown each model’s flag rate since August 6, 2026.

A blocker is a criterion a maintainer would refuse to merge over, so one that’s too strict rejects valid patches, an error in the grader rather than the model. Cognition audited more than 1,000 criteria to find the 75. Diamond, the 50 hardest of the 150 tasks, no longer held the 50 hardest after those changes and was too noisy at its low solve rates, so Cognition now reports only Main (100 tasks) and Extended (150). Per the 1.1 post, absolute scores moved but model rankings didn’t substantially change from 1.0.