KernelGenBench-MS (Claude Code): leaderboard
Metric: Clean pass rate (%; share of the 210 KernelGenBench-MS operators (110 PyTorch ATen, 50 vLLM, 50 cuBLAS) whose generated Triton kernel passes numerical validation on every test case of its combinatorial shape, dtype and layout suite and every anti-hack check, as a drop-in replacement; the Claude Code agent with execution feedback; NVIDIA A100; 30-minute budget per operator). Source: arxiv.org. Saturation forecast: Around December 2026. 4 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.6 | 87 |
| 2 | GLM-5 | 67 |
| 3 | Qwen 3.5 27B | 62 |
| 4 | MiniMax-M2.5 | 49 |
Interactive version: theaggregate.ai/benchmark?slug=kernelgenbench-ms-claude-code · How It Works · Data refreshed daily, snapshot 2026-09-29.