sudo L7
Code
Surge AI's 60-task evaluation of engineering judgment: agents navigate unfamiliar systems, choose architectures, identify hidden risks, and assess whether technically correct code meets the actual request. Published aggregate percentages are attributed to the agent harness Surge names for each row; no single-run pass count is inferred.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 41.7 |
| Claude Sonnet 5.5 | 41.7 |
| GPT-6 Astra | 33.9 |
| Claude Fable 5.1 | 31.1 |
| GPT-6.1 Sol | 28.9 |
| Gemini 4 Argon | 18.9 |
| Grok 4.7 | 15.6 |
| Muse Wonder | 15.6 |
| GLM-5.3 | 14.4 |
| Muse Spark 1.3 | 13.9 |
| Kimi K3 | 13.3 |
| Gemini 3.8 Flash | 11.7 |
| Mistral Large 4 Preview | 11.7 |