NVIDIA ComputeEval: leaderboard
CUDA programming problems from NVIDIA, 127 at launch and 566 in the 2026.1 release, covering kernels, runtime APIs, CCCL, cuBLAS, GPU math libraries, cuDNN and cuTile in C++ and Python; pass@1 on held-out functional tests.
Metric: Pass@1 (%, zero-shot, 2026.1 release, 566 problems). Source: github.com. Source: huggingface.co. Saturation forecast: Around January 2027. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 (High) | 70.7 | #63 (Claude Opus 4.6) |
| 2 | Claude Opus 4.6 (Medium) | 68.6 | #63 (Claude Opus 4.6) |
| 3 | GPT-5.4 (High) | 65.4 | #79 (GPT-5.4) |
| 4 | Claude Sonnet 4.6 (High) | 65 | #90 (Claude Sonnet 4.6) |
| 5 | GPT-5.4 (Medium) | 64.8 | #79 (GPT-5.4) |
| 6 | Claude Sonnet 4.6 (Medium) | 62.9 | #90 (Claude Sonnet 4.6) |
| 7 | Kimi K2.5 (Thinking) | 53.5 | #144 (Kimi K2.5) |
| 8 | GPT-5 Mini (High) | 53 | #186 (GPT-5 Mini) |
| 9 | GLM-5 (Thinking) | 51.4 | #141 (GLM-5) |
| 10 | GPT-OSS-120B (High) | 50.7 | #341 (GPT-OSS-120B) |
| 11 | GPT-5 Mini (Medium) | 49.6 | #186 (GPT-5 Mini) |
| 12 | Gemini 3.1 Flash Lite (High) | 48.9 | #191 (Gemini 3.1 Flash Lite) |
| 13 | Gemini 3.1 Flash Lite (Medium) | 48.4 | #191 (Gemini 3.1 Flash Lite) |
| 14 | GPT-OSS-120B (Medium) | 47.3 | #341 (GPT-OSS-120B) |
| 15 | GPT-OSS-20B (Medium) | 35.7 | #518 (GPT-OSS-20B) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=nvidia-computeeval · How It Works · Data refreshed daily, snapshot 2026-10-20.