FrontierCode 2.0 Leaderboard
Cognition's 132-task benchmark asks whether an AI-generated change is mergeable—not merely whether it passes tests. Results below are the best reported model-and-harness configurations on Cognition's live board.
Last checked: October 30, 2026 · Includes Gemini 4 Carbon (40.0%, $4.04 per rollout) and Gemini 4 Argon (26.3%, $6.51 per rollout).
| # | Model · best effort | Provider | Main score | Pass rate | Flag rate | Cost / rollout | Output tokens | Harness |
|---|
Cognition defines score as the weighted aggregate of rubric items; attempts failing a blocking criterion score zero. Rollout costs are mean USD at list API pricing (including public discounts), and are separate from token-rate pricing. Flag rate is the fraction of runs flagged for unfair internet use.
FrontierCode 2.0 score comparison
How to read this table
FrontierCode 2.0 evaluates changes in real repositories against maintainer-authored criteria covering correctness, tests, scope discipline, style, and project conventions. Cognition restricts solution-bearing internet access: flagged attempts receive zero. The number at left is Cognition's order by the model's strongest reported effort; ties and cross-harness comparisons should not be interpreted as controlled head-to-heads.
API prices are not shown as though they were benchmark costs: the listed spend is the actual mean per rollout under Cognition's evaluation setup. For example, GPT-6.1 Sol's 39.0% cost $1.03 per rollout, while Claude Opus 5.5 led at 45.5% for $12.82. Different harnesses, task consumption, and model prices affect those totals.
- Cognition live FrontierCode leaderboard — scores, harnesses, costs, output tokens, and changelog.
- Cognition: FrontierCode 2.0 benchmark and methodology.
- Cognition: FrontierCode 1.1 archive.