Skip to content
All systems operational0 AI providers monitored, polled every 2 minutes
Live status
All benchmarks

FrontierCode v1.1 leaderboard

FrontierCode asks a harder question than most coding benchmarks: not whether a patch passes tests, but whether a maintainer would merge it. Cognition built it with roughly 36 open-source maintainers from projects including Celery, Budibase, uppy, and Mattermost, who authored tasks against real issues in their own repositories. Grading combines held-out tests with a maintainer-written rubric covering behavioral correctness, regression safety, test quality, scope discipline, build and lint cleanliness, and adherence to project conventions. The Main split is the 100 hardest tasks; an Extended split covers all 150.

Current leader
Claude Fable 5(Anthropic)53.5%

Last refreshed 2026-08-26. 3 models scored on this benchmark.

Full leaderboard

#ModelProviderScoreReleased
1Claude Fable 5Anthropic53.5%2026-06
2Claude Opus 5Anthropic53.4%vendor2026-07
3Claude Opus 4.8Anthropic46.5%2026-05

Score interpretation

TensorFeed reports the Main split. Two things to watch. Scores vary substantially by reasoning effort, and the leaderboard publishes a separate figure per effort level, so a headline number is meaningless without knowing which one it came from. And the benchmark is deliberately unpublished, with no public repo, which means aggregator sites quoting FrontierCode figures for models that are not on Cognition's own board are reporting something other than a FrontierCode run. Check the maintainer board before trusting a number.

50%+
Frontier. Produces mergeable work on the hardest tasks about half the time.
35-50%
Strong. Real contributions, but review is still mandatory.
20-35%
Useful drafts. Expect substantial rework before merge.
< 20%
Output is a starting point, not a contribution.

Why this matters for AI agents

Test-passing and mergeable are different bars, and the gap between them is where most coding-agent disappointment lives. A patch that turns tests green while sprawling across unrelated files, ignoring project conventions, or quietly weakening a test is a patch a human has to redo. This is the only benchmark we track that puts a maintainer's judgment in the scoring loop, which makes it the best available signal for whether agent output will survive code review.

Other benchmarks

Premium API: time-series for FrontierCode v1.1

The leaderboard above is a snapshot. Want to see how a model's FrontierCode v1.1 score has moved over the last 30-90 days, or set a webhook that fires when a score crosses a threshold? The premium API has both:

FrontierCode v1.1 source ·Last refreshed 2026-08-26·Max score 100