What does a FrontierCode score mean?

A FrontierCode score is the weighted share of maintainer-written criteria an agent’s patches meet, with any patch that fails a merge-blocking criterion scored zero, so it never exceeds the pass rate, the share of patches clearing every blocker. On Cognition’s FrontierCode 1.1 Main leaderboard, as of October 18, 2026, Anthropic’s Claude Opus 5.5 scores 54.6% with a 59.6% pass rate.

Each task carries two kinds of criteria. Blockers are the hard stops of a code review: correctness, regressions, and touching files the patch shouldn’t. Non-blockers are quality signals like style, type safety, and readability. A patch that clears every blocker then earns the weighted total of everything it meets, so a working patch can still lose credit, or fail outright, for unneeded refactors, missing tests, or ignoring the repo’s conventions. Some of those criteria are graded by an LLM reviewer against the maintainer’s written standard, which makes part of the score a rubric in the LLM-as-judge sense.

The headline number is also tied to a subset and a setting. Main is the 100 hardest of FrontierCode’s 150 tasks, and the same Opus 5.5 configuration scores 65.3% on the full Extended set, per Cognition’s leaderboard data. Cognition’s launch post says each model runs five times at every reasoning effort, and the leaderboard reports the average at whichever effort scored best, so the listed score comes from each model’s best setting, which for Opus 5.5 is medium rather than max.