FrontierCode 2.0 - Score
Code · 2026-10-27
FrontierCode 2.0 weighted score across Cognition's rebuilt suite of 132 production-code tasks written by open-source maintainers. Compared with 1.1 it draws on more repositories and task categories, including security patches and performance optimizations, adds blocking criteria with emphasis on thoroughness, loosens scope criteria, and requires larger changes. A trial that fails any blocker criterion or is flagged for consulting solution-bearing sources receives zero. Scores are not comparable with FrontierCode 1.x.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 45.5 |
| GPT-6 Astra | 42.8 |
| Claude Sonnet 5.5 | 41.0 |
| GPT-6.1 Sol | 39.0 |
| GPT-6 Sol | 34.3 |
| Claude Fable 5.1 | 32.4 |
| Grok 4.7 | 28.6 |
| Claude Haiku 5.5 | 27.7 |
| GPT-5.6 Sol | 27.6 |
| Claude Opus 5 | 26.7 |
| Gemini 4 Argon | 26.3 |
| Muse Wonder | 24.3 |
| SWE-2 | 23.8 |
| GPT-5.5 | 22.6 |
| GPT-6 Luna | 21.9 |
| GLM-5.3 | 17.8 |
| Kimi K3 | 17.7 |
| Opus 4.8 | 14.7 |
| GLM-5.3 Flash | 14.2 |
| Inkling-Small | 3.2 |
| Beam 1.0 501B A23B | 2.9 |