ProgramBench: leaderboard
Meta and Stanford benchmark testing whether language-model agents can rebuild complete programs from only a compiled binary and documentation. Agents use mini-SWE-agent across 200 open-source program recreation tasks and are scored by hidden behavioral tests.
Metric: Resolved (%). Source: programbench.com. Status: years away from saturation. 21 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 5 (xHigh) | 4.5 |
| 2 | GPT-5.6 Sol (xHigh) | 1 |
| 3 | GPT-5.6 Sol | 0.5 |
| 4 | GPT-5.5 (xHigh) | 0.5 |
| 5 | Gemini 3.6 Flash | 0.5 |
| 6 | GPT-5.5 (High) | 0.5 |
| 7 | Claude Sonnet 4.6 | 0 |
| 8 | Gemini 3.1 Pro (Preview) | 0 |
| 9 | GPT-5.5 | 0 |
| 10 | GPT-5.4 | 0 |
| 11 | GPT-5 Mini | 0 |
| 12 | Claude Opus 4.7 | 0 |
| 13 | Claude Opus 4.6 | 0 |
| 14 | Claude Haiku 4.5 | 0 |
| 15 | Gemini 3 Flash | 0 |
Interactive version: theaggregate.ai/benchmark?slug=programbench · How It Works · Data refreshed daily, snapshot 2026-09-05.