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. Saturation forecast: Around 2031. 25 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 (xHigh) | 4.5 | #12 (Claude Opus 5) |
| 2 | Muse Spark 1.3 (Max) | 2.5 | #24 (Muse Spark 1.3) |
| 3 | GPT-5.6 Sol (xHigh) | 1 | #19 (GPT-5.6 Sol) |
| 4 | Muse Spark 1.3 (xHigh) | 1 | #24 (Muse Spark 1.3) |
| 5 | GPT-5.6 Sol | 0.5 | #19 |
| 6 | Gemini 3.6 Flash | 0.5 | #51 |
| 7 | GPT-5.5 (xHigh) | 0.5 | #31 (GPT-5.5) |
| 8 | GPT-5.5 (High) | 0.5 | #31 (GPT-5.5) |
| 9 | Gemini 3.1 Pro (Preview) | 0 | #58 |
| 10 | Claude Sonnet 4.6 | 0 | #90 |
| 11 | GPT-5.4 | 0 | #79 |
| 12 | GPT-5.5 | 0 | #31 |
| 13 | Claude Opus 4.6 | 0 | #63 |
| 14 | Claude Haiku 4.5 | 0 | #280 |
| 15 | Gemini 3 Flash | 0 | #99 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=programbench · How It Works · Data refreshed daily, snapshot 2026-10-20.