KernelBench CUDA: leaderboard

Metric: Mean Peak Fraction Across Problems (ratio). Source: benchmarklist.com. 19 models tracked.

Top models

#ModelScore
1Claude Opus 5 (Max)0.79
2Claude Opus 4.8 (Max)0.37
3Qwen 3.8 Max (xHigh)0.28
4Claude Fable 50.26
5Grok 4.5 (High)0.18
6Kimi K30.18
7DeepSeek V4 Pro0.12
8GLM-5.30.1
9Grok 4.6 (xHigh)0.08
10DeepSeek V4 Flash (0731)0.07
11GLM-5.3 Flash0.04
12Grok 4.50.03

Interactive version: theaggregate.ai/benchmark?slug=kernelbench-cuda · How It Works · Data refreshed daily, snapshot 2026-09-19.