ProgramBench (Mercor 100-task subset)
Code
Mercor's 100-task subset of ProgramBench's 200 reverse-engineering tasks, in which an agent rebuilds a program from its executable and documentation. A task resolves only if every hidden behavioral test passes.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 20.0 |
| Claude Fable 5.1 | 11.3 |
| GPT-6 Astra | 8.0 |
| GPT-6.1 Sol | 6.3 |
| Claude Opus 5 | 4.3 |
| Claude Haiku 5.5 | 3.0 |
| GPT-6 Sol | 2.3 |
| Grok 4.6 | 1.7 |
| Gemini 3.7 Flash | 1.0 |
| Gemini 3.8 Flash | 1.0 |
| GPT-5.5 | 1.0 |
| GPT-5.6 Sol | 1.0 |
| GPT-5.6 Terra | 1.0 |
| GPT-6 Luna | 1.0 |
| Sonnet 5 | 0.3 |
| DeepSeek V4 Pro 0813 | 0.3 |
| Gemini 3.6 Flash | 0.3 |
| GLM-5.2 | 0.3 |
| Claude Fable 5 | 0.0 |
| GPT-5.6 Luna | 0.0 |
| Grok 4.7 | 0.0 |
| Inkling | 0.0 |
| MiniMax M3 | 0.0 |