Benchmarks · coding
KernelBench CUDA (kernelbench.com)
Coding agents write four CUDA-only GPU kernels (GLM-5.2 fused MoE, DeepSeek sparse attention, MegaQwen decode, grid MinGRU scan) where Triton and other DSLs fail, one unlimited session per problem on an RTX PRO 6000, each run audited for reward hacking. Mean per-problem score over the full deck (peak fraction of roofline; sparse attention against a reference latency, so it can exceed 100%); failed, flagged or unaudited problems count 0. Independent of Stanford's KernelBench.
22 models · 22 results · updated 2026-10-18 · kernelbench.com ↗
Leaderboard
Sonnet 5.5 max
More from this benchmark
7 published measuresScore vs any published measure7 measures: sub-scores, splits, counts
Breakdown
darker is better within each column · click a column to sort| Model | ||||
|---|---|---|---|---|
| Sonnet 5.5 max, fallback | 0.113 | 1.36 | 0.08 | 0.944 |
| Opus 5.5 xhigh, fallback | 0.104 | 1.1 | 0.0739 | 0.794 |
| Opus 5 max | 0.0655 | 1.96 | ||
| Gemini 4 Argon high | 0.0992 | 0.518 | 0.054 | 0.422 |
| Fable 5 max | 0.0804 | 0.727 | 0.0578 | 0.191 |
| Grok 4.7 xhigh | 0.0948 | 0.1 | 0.0455 | 0.64 |
| Fable 5.1 max, fallback | 0.102 | 0.0643 | 0.709 | |
| GPT-6 Astra Pro xhigh | 0.0945 | 0.049 | 0.684 | |
| GPT-6 Sol xhigh | 0.0936 | 0.0376 | 0.0398 | 0.645 |
| GPT-6.1 Sol xhigh | 0.0941 | 0.352 | 0.0386 | 0.305 |
| DeepSeek V4.1 Flash 1017 max | 0.0965 | 0.37 | 0.0812 | 0.2 |
| GPT-6 Luna xhigh | 0.0798 | 0.0377 | 0.0286 | 0.537 |
| DeepSeek V4.1 Flash max | 0.0946 | 0.502 | 0.0539 | |
| Haiku 5.5 max | 0.0947 | 0.1 | 0.0458 | 0.34 |
| Gemini 3.8 Flash high | 0.0958 | 0.0426 | 0.364 | |
| Qwen 3.8 Max xhigh | 0.101 | 0.285 | ||
| Kimi K3 max | 0.0595 | 0.047 | 0.174 | |
| DeepSeek V4 Flash 0731 max | 0.0411 | 0.0281 | 0.145 | |
| Grok 4.6 xhigh | 0.0939 | 0.0542 | ||
| Muse Spark 1.3 max | 0.0879 | 0.0251 | 0.0333 | |
| GLM 5.3 max | 0.0997 | |||
| GLM 5.3 Flash max | 0.0445 | 0.0392 |
All results
22 rows| # | Model | Effort | Score | Harness |
|---|---|---|---|---|
| 1 | Sonnet 5.5 | max | 62.5% | claude |
| 2 | Opus 5.5 | xhigh | 51.7% | claude |
| 3 | Opus 5 | max | 50.7% | or-opus |
| 4 | Gemini 4 Argon | high | 27.3% | agy |
| 5 | Fable 5 | max | 26.4% | or-fable |
| 6 | Grok 4.7 | xhigh | 22.0% | grok |
| 7 | Fable 5.1 | max | 21.9% | or-fable |
| 8 | GPT-6 Astra Pro | xhigh | 20.7% | codex |
| 9 | GPT-6 Sol | xhigh | 20.4% | codex |
| 10 | GPT-6.1 Sol | xhigh | 19.8% | codex |
| 11 | DeepSeek V4.1 Flash 1017 | max | 18.7% | or-fable |
| 12 | GPT-6 Luna | xhigh | 17.1% | codex |
| 13 | DeepSeek V4.1 Flash | max | 16.3% | deepseek-claude |
| 14 | Haiku 5.5 | max | 14.5% | claude |
| 15 | Gemini 3.8 Flash | high | 12.6% | agy |
| 16 | Qwen 3.8 Max | xhigh | 9.6% | or-fable |
| 17 | Kimi K3 | max | 7.0% | kinetic-claude |
| 18 | DeepSeek V4 Flash 0731 | max | 5.4% | or-fable |
| 19 | Grok 4.6 | xhigh | 3.7% | grok |
| 20 | Muse Spark 1.3 | max | 3.7% | muse |
| 21 | GLM 5.3 | max | 2.5% | zai-claude |
| 22 | GLM 5.3 Flash | max | 2.1% | or-fable |