CUDABench (Level 2) - Functional Correctness: leaderboard
Metric: Functional consistency pass@1 (%): share of kernels that compile and reproduce the reference outputs on every input scale; CUDABench-Set: 500 text-to-CUDA tasks across application domains, each with five input scales, Level 2 prompts (task name and algorithm description without hardware guidance); kernels compiled and run on an NVIDIA A40, pass@1; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 82.8 | #138 |
| 2 | Gemini 3 Flash (Preview) | 77.6 | #78 |
| 3 | GPT-5.2 (2025-12-11) (High) | 75.2 | |
| 4 | Qwen 3 Max (2025-09-23) | 73.8 | #249 |
| 5 | DeepSeek V3.2 (Thinking) | 63.4 | #198 (DeepSeek V3.2) |
| 6 | MiniMax-M2.1 | 57.6 | #283 |
| 7 | Kimi K2 (Thinking) | 53.8 | #236 (Kimi K2) |
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-2-functional-correctness · How It Works · Data refreshed daily, snapshot 2026-10-11.