# ProgramBench: Can Language Models Rebuild Programs From Scratch? (ProgramBench)

> A cleanroom software-engineering benchmark where agents receive only a compiled executable and documentation, then must architect and implement a complete codebase that reproduces the original program's behavior.

Canonical page: https://benchlm.ai/benchmarks/programbench

- Category: [Coding](/coding)
- Last updated: Sep. 9, 2026

## About ProgramBench

- Year: 2026
- Tasks: 200 program reconstruction tasks
- Format: Cleanroom executable reimplementation
- Difficulty: Full-repository software architecture
- Paper: [ProgramBench: Can Language Models Rebuild Programs From Scratch?](https://programbench.com/static/paper.pdf)

ProgramBench turns open-source projects into cleanroom reconstruction tasks. Each task starts from an execute-only binary and usage documentation, with no source code, internet, decompilation, or prescribed skeleton. Evaluation uses hidden behavioral tests generated through agent-driven fuzzing. BenchLM shows ProgramBench as display-only because all current public rows are tied at 0% fully resolved and the visible score is the auxiliary almost-resolved metric.

ProgramBench is currently displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Leaderboard (21 models)

| Rank | Model | Creator | Score |
|------|-------|---------|-------|
| 1 | [Claude Opus 5](/models/claude-opus-5) | Anthropic | 37.0% |
| 2 | [Claude Opus 4.8](/models/claude-opus-4-8) | Anthropic | 16.5% |
| 3 | [GPT-5.6 Sol](/models/gpt-5-6-sol) | OpenAI | 15.5% |
| 4 | [GPT-5.5](/models/gpt-5-5) | OpenAI | 13.5% |
| 5 | [GLM-5.2](/models/glm-5-2) | Z.AI | 8.5% |
| 6 | [Gemini 3.7 Flash](/models/gemini-3-7-flash) | Google | 5.5% |
| 7 | [GPT-5.5](/models/gpt-5-5) | OpenAI | 5.0% |
| 8 | [Claude Opus 4.7](/models/claude-opus-4-7) | Anthropic | 4.5% |
| 9 | [Gemini 3.6 Flash](/models/gemini-3-6-flash) | Google | 4.0% |
| 10 | [Gemini 3.5 Flash](/models/gemini-3-5-flash) | Google | 3.0% |
| 11 | [Claude Opus 4.7](/models/claude-opus-4-7) | Anthropic | 3.0% |
| 12 | [GPT-5.6 Sol](/models/gpt-5-6-sol) | OpenAI | 2.5% |
| 13 | [Claude Opus 4.6](/models/claude-opus-4-6) | Anthropic | 2.5% |
| 14 | [GPT-5.5](/models/gpt-5-5) | OpenAI | 1.5% |
| 15 | [Claude Sonnet 4.6](/models/claude-sonnet-4-6) | Anthropic | 1.0% |
| 16 | [GPT-5.4](/models/gpt-5-4) | OpenAI | 0.0% |
| 17 | [Gemini 3.1 Pro](/models/gemini-3-1-pro) | Google | 0.0% |
| 18 | [Gemini 3 Flash](/models/gemini-3-flash) | Google | 0.0% |
| 19 | [Claude Haiku 4.5](/models/claude-haiku-4-5) | Anthropic | 0.0% |
| 20 | [GPT-5.4 mini](/models/gpt-5-4-mini) | OpenAI | 0.0% |
| 21 | [GPT-5 mini](/models/gpt-5-mini) | OpenAI | 0.0% |

## FAQ

### What does ProgramBench measure?

A cleanroom software-engineering benchmark where agents receive only a compiled executable and documentation, then must architect and implement a complete codebase that reproduces the original program's behavior.

### Which model leads the published ProgramBench snapshot?

Claude Opus 5 currently leads the published ProgramBench snapshot with a score of 37.0%.

### How many models are evaluated on ProgramBench?

The Sep. 9, 2026 contains 21 AI models.

### Does ProgramBench affect BenchLM's overall score?

Not directly. ProgramBench is still displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.