CVDP (517-problem set)
Code
NVIDIA's Comprehensive Verilog Design Problems benchmark as Mercor runs it: 517 RTL design and verification problems covering spec-to-RTL, code completion and modification, debugging, testbench generation and RTL comprehension. Pass@1 is the share of problems whose checks pass.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 64.1 |
| Claude Fable 5.1 | 62.6 |
| Claude Opus 5 | 62.2 |
| Claude Sonnet 5.5 | 60.6 |
| Grok 4.6 | 58.0 |
| Claude Haiku 5.5 | 57.0 |
| Sonnet 5 | 56.4 |
| Kimi K3 | 55.5 |
| GPT-5.6 Sol | 55.4 |
| GPT-6.1 Sol | 55.1 |
| GPT-6 Astra | 54.9 |
| Grok 4.7 | 54.9 |
| Muse Spark 1.3 | 54.3 |
| DeepSeek V4 Pro 0813 | 53.1 |
| GLM-5.3 | 53.1 |
| GPT-6 Sol | 52.7 |
| Gemini 3.1 Pro Preview | 51.3 |
| GPT-5.6 Terra | 50.9 |
| GPT-6 Luna | 50.8 |
| Gemini 3.8 Flash | 50.7 |
| GPT-5.6 Luna | 48.4 |
| MiniMax M3 | 44.8 |