FrontierCode 1.1 Main (Artificial Analysis)
Code
Artificial Analysis' implementation of Cognition FrontierCode 1.1 Main on 100 private repository tasks. Stirrup runs three attempts per task with a 250-turn limit and GPT-6 Luna medium as judge. Blocking failures and solution-lookup flags score zero; other rubric checks contribute to a mean score, averaged equally over tasks and attempts. This evaluator protocol is kept separate from Cognition's native runs.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 57.1 |
| GPT-6.1 Sol | 49.1 |
| Claude Fable 5.1 | 49.0 |
| Gemini 4 Carbon | 48.5 |
| Kimi K3.1 Pro | 47.1 |
| GPT-6 Astra | 47.0 |
| Muse Wonder | 46.5 |
| Claude Sonnet 5.5 | 46.1 |
| Claude Haiku 5.5 | 45.5 |
| MiMo-V2.6-Pro | 45.4 |
| Muse Wonder Full | 44.9 |
| DeepSeek V4.1 Flash | 44.0 |
| GPT-6.1 Terra | 43.7 |
| DeepSeek V4.1 Flash 1017 | 43.7 |
| Kimi K3 | 43.6 |
| Gemini 4 Argon | 43.5 |
| Qwen3.8 Max (0902) | 43.2 |
| MiMo-V2.6-Flash | 42.1 |
| GPT-6 Luna | 41.9 |
| Mistral Large 4 Preview | 41.5 |
| GLM-5.3 Flash | 41.0 |
| Muse Spark 1.3 | 40.6 |
| GLM-5.3 | 40.1 |
| Grok 4.7 | 39.6 |
| DeepSeek V4 Pro 0813 | 38.8 |
| MiniMax M3.1 Flash | 38.6 |
| Step 5 Preview | 38.4 |
| Gemini 3.8 Flash | 37.1 |
| Qwen3.8 27B | 34.7 |
| K2 Horizon 375B A23B | 32.1 |
| Muse Spark Open | 27.6 |
| Nemotron 3.5 Super VL 120B A12B (precision unspecified) | 25.5 |
| Beam 1.0 501B A23B | 24.3 |
| MiniMax M3 | 23.0 |
| Inkling | 15.8 |
| Muse Glimmer 30B | 13.9 |
| Nemotron 3 Ultra 550B A55B | 10.5 |
| gpt-oss-120b | 1.2 |
| Command A+ (May 2026, BF16) | 1.2 |
| Inkling-Small | 0.6 |