KernelBench CUDA Normalized (RTX PRO 6000)
Code · 2026-09-30
KernelBench CUDA RTX PRO 6000 board: mean audited valid cell score relative to the board-best score per problem, scaled to 0–100 over the full four-problem deck. Missing/invalid cells contribute zero. Best published cells may come from different harness routes and reasoning settings of the same model; distinct from the earlier raw roofline metric.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Sonnet 5.5 | 87.0 |
| Claude Opus 5.5 | 76.3 |
| Claude Fable 5.1 | 51.6 |
| Claude Fable 5 | 51.6 |
| DeepSeek V4.1 Flash | 50.6 |
| Claude Opus 5 | 45.5 |
| Grok 4.7 | 45.2 |
| GPT-6 Astra Pro | 44.9 |
| GPT-6 Sol | 42.0 |
| Grok 4.6 | 37.7 |
| GPT-6 Luna | 34.1 |
| Muse Spark 1.3 | 30.3 |
| Kimi K3 | 30.1 |
| Grok 4.5 | 29.5 |
| Opus 4.8 | 24.9 |
| GLM-5.3 | 22.1 |
| DeepSeek-V4-Flash-0731 | 19.7 |
| Gemini 3.8 Flash | 19.7 |