KernelBench Hub - CUDA: leaderboard
kernelbench.com CUDA suite (independent of Stanford KernelBench): agentic CUDA kernel optimization scored as percent of hardware roofline achieved, best across RTX PRO 6000, H100 and B200 runs.
Metric: Best % of Hardware Roofline. Source: kernelbench.com. Saturation forecast: Around May 2028. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 | 196.1 | #12 |
| 2 | Claude Sonnet 5.5 | 94.42 | #11 |
| 3 | Claude Opus 5.5 | 79.38 | #4 |
| 4 | Claude Fable 5.1 | 70.93 | #9 |
| 5 | GPT-6 Sol | 64.47 | #15 |
| 6 | Grok 4.7 | 63.98 | #26 |
| 7 | GPT-6 Luna | 53.69 | #60 |
| 8 | Gemini 4 Argon (High) | 42.2 | #8 (Gemini 4 Argon) |
| 9 | Gemini 3.8 Flash (High) | 36.37 | #23 (Gemini 3.8 Flash) |
| 10 | Claude Haiku 5.5 | 34.02 | #37 |
| 11 | GPT-6.1 Sol | 30.52 | #7 |
| 12 | Qwen 3.8 Max | 28.48 | #32 |
| 13 | DeepSeek V4.1 Flash (1017) | 20.02 | #20 |
| 14 | Claude Fable 5 | 19.09 | #16 |
| 15 | Kimi K3 | 17.38 | #30 |
No result here: #6 GPT-6 Astra, #10 Muse Wonder Full.
Interactive version: theaggregate.ai/benchmark?slug=kernelbench-hub-cuda · How It Works · Data refreshed daily, snapshot 2026-10-23.