KernelBench-Verified - Level 1: leaderboard
LLMs write custom CUDA kernels for the 100 single-operator KernelBench problems (matrix multiplication, softmax, normalization, activations), timed against PyTorch with TF32 tensor cores enabled after correctness checks on four hidden input distributions; from KernelBench-Verified (2026). Fast@1, the share of problems where the best correct kernel beats PyTorch, in percent; higher is better.
Metric: Fast@1 (%; 100 single-operator problems; share of problems whose fastest correct sample beats the baseline; single-turn KernelBench prompt, best of five samples; correctness gated on four hidden input distributions (original, x3, x0.01, negated); timing against a TF32-enabled eager PyTorch baseline on an H200). Source: arxiv.org. Saturation forecast: Around March 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 52 | #49 |
| 2 | Gemini 3.1 Pro (Preview) | 47 | #94 |
| 3 | Claude Sonnet 4.6 | 36 | #161 |
| 4 | Claude Opus 4.7 | 35 | #79 |
| 5 | Gemini 3 Flash (Preview) | 35 | #144 |
| 6 | Kimi K2.6 | 34 | #178 |
| 7 | Claude Opus 4.8 | 34 | #56 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=kernelbench-verified-level-1 · How It Works · Data refreshed daily, snapshot 2026-11-13.