KernelBench CUDA: leaderboard
KernelBench CUDA is a text-only coding benchmark for GPU kernel tasks, including fused MoE, sparse attention, decode, and simulation kernels. It evaluates agents on CUDA-only kernel implementation. The pinned metric, Mean Peak Fraction Across Problems, measures performance as a ratio relative to the best achievable peak across problems. Source: BenchmarkList’s attributed result table.
Metric: Mean Peak Fraction Across Problems (ratio). Source: benchmarklist.com. Saturation forecast: Around January 2028. 33 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 5.5 (Max) | 0.62 | #14 (Claude Sonnet 5.5) |
| 2 | Claude Opus 5.5 (xHigh) | 0.52 | #3 (Claude Opus 5.5) |
| 3 | Claude Opus 5 (Max) | 0.51 | #16 (Claude Opus 5) |
| 4 | Claude Opus 4.8 (Max) | 0.37 | #39 (Claude Opus 4.8) |
| 5 | Gemini 4 Argon (High) | 0.27 | #9 (Gemini 4 Argon) |
| 6 | Claude Fable 5 | 0.26 | #20 |
| 7 | Grok 4.7 (xHigh) | 0.22 | #26 (Grok 4.7) |
| 8 | Claude Fable 5.1 (Max) | 0.22 | #10 (Claude Fable 5.1) |
| 9 | GPT-6 Sol (xHigh) | 0.2 | #19 (GPT-6 Sol) |
| 10 | GPT-6.1 Sol (xHigh) | 0.2 | #7 (GPT-6.1 Sol) |
| 11 | DeepSeek V4.1 Flash (1017) | 0.19 | #27 |
| 12 | Grok 4.5 (High) | 0.18 | #54 (Grok 4.5) |
| 13 | GPT-6 Luna (xHigh) | 0.17 | #61 (GPT-6 Luna) |
| 14 | Claude Haiku 5.5 (Max) | 0.15 | #43 (Claude Haiku 5.5) |
| 15 | Gemini 3.8 Flash (High) | 0.13 | #30 (Gemini 3.8 Flash) |
No result here: #6 GPT-6 Astra.
Interactive version: theaggregate.ai/benchmark?slug=kernelbench-cuda · How It Works · Data refreshed daily, snapshot 2026-11-05.