CUDABench (Level 1): leaderboard
Metric: CUDABench-Score pass@1 (%): mean over tasks of the roofline Performance-Score (achieved over attainable GFLOP/s) for kernels that compile and match the reference outputs, zero otherwise; CUDABench-Set: 500 text-to-CUDA tasks across application domains, each with five input scales, Level 1 prompts (task name, algorithm description and explicit CUDA implementation guidelines); kernels compiled and run on an NVIDIA A40, pass@1; higher is better. Source: arxiv.org. Saturation forecast: Around December 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 (2025-12-11) (High) | 40.9 | |
| 2 | Claude Sonnet 4.5 | 40.2 | #138 |
| 3 | Gemini 3 Flash (Preview) | 40.1 | #78 |
| 4 | Kimi K2 (Thinking) | 35.1 | #236 (Kimi K2) |
| 5 | Qwen 3 Max (2025-09-23) | 32.1 | #249 |
| 6 | DeepSeek V3.2 (Thinking) | 31.6 | #198 (DeepSeek V3.2) |
| 7 | MiniMax-M2.1 | 26.7 | #283 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=cudabench-level-1 · How It Works · Data refreshed daily, snapshot 2026-10-11.