# Vals ProgramBench (ProgramBench)

> Can language models rebuild programs from scratch?

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

- Category: [Coding](/coding)
- Last updated: September 27, 2026

## About ProgramBench

- Year: 2026
- Tasks: Program reconstruction tasks
- Format: Accuracy score
- Difficulty: Cleanroom software engineering
- Paper: [ProgramBench](https://www.vals.ai/benchmarks/programbench)

BenchLM mirrors the public Vals AI ProgramBench leaderboard as display-only external evidence. The captured snapshot preserves overall scores, task-level scores where Vals publishes them, uncertainty, latency, and cost-per-test metadata. It is excluded from BenchLM weighted rankings.

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

## Leaderboard (53 models)

| Rank | Model | Configuration | Creator | Score |
|------|-------|---------------|---------|-------|
| 1 | [Claude Opus 5.5](/models/claude-opus-5-5) | — | Anthropic | 18.50% |
| 2 | [Claude Fable 5.1](/models/claude-fable-5-1) | — | Anthropic | 7.00% |
| 3 | [GPT-6 Astra](/models/gpt-6-astra) | max reasoning | OpenAI | 5.50% |
| 4 | [Claude Opus 5](/models/claude-opus-5) | — | Anthropic | 3.00% |
| 5 | [Claude Fable 5](/models/claude-fable) | — | Anthropic | 2.00% |
| 6 | [GPT-6 Sol](/models/gpt-6-sol) | max reasoning | OpenAI | 2.00% |
| 7 | [Kimi K3](/models/kimi-k3) | — | Moonshot AI | 2.00% |
| 8 | [GPT-5.6 Sol](/models/gpt-5-6-sol) | max reasoning | OpenAI | 1.50% |
| 9 | [GLM-5.3](/models/glm-5-3) | max reasoning | Z.AI | 1.50% |
| 10 | [Claude Opus 4.8](/models/claude-opus-4-8) | — | Anthropic | 1.00% |
| 11 | [Gemini 3.8 Flash](/models/gemini-3-8-flash) | high reasoning | Google | 1.00% |
| 12 | [Grok 4.7](/models/grok-4-7) | xhigh reasoning | xAI | 0.50% |
| 13 | [MiMo-V2.6-Flash](/models/mimo-v2-6-flash) | — | Xiaomi | 0.50% |
| 14 | [GPT-5.6 Terra](/models/gpt-5-6-terra) | max reasoning | OpenAI | 0.50% |
| 15 | [MiMo-V2.6-Pro](/models/mimo-v2-6-pro) | — | Xiaomi | 0.50% |
| 16 | [GPT-5.5](/models/gpt-5-5) | xhigh reasoning | OpenAI | 0.50% |
| 17 | [GLM-5.2](/models/glm-5-2) | max reasoning | Z.AI | 0.50% |
| 18 | [GPT-6 Luna](/models/gpt-6-luna) | max reasoning | OpenAI | 0.50% |
| 19 | [GPT-5.4 2026 03 05 High](https://www.vals.ai/models/openai_gpt-5.4-2026-03-05-high) | high reasoning | OpenAI | 0.50% |
| 20 | [Claude Sonnet 4.6](/models/claude-sonnet-4-6) | — | Anthropic | 0.50% |
| 21 | [Inkling-Small](/models/inkling-small) | 0.99 reasoning | Thinking Machines Lab | 0.50% |
| 22 | [Hy4 preview](/models/hy4-preview) | — | Tencent | 0.00% |
| 23 | [DeepSeek V4 Pro 0813](/models/deepseek-v4-pro-0813) | max reasoning | DeepSeek | 0.00% |
| 24 | [Claude Sonnet 5](/models/claude-sonnet-5) | — | Anthropic | 0.00% |
| 25 | [Qwen3.8 Max](/models/qwen3-8-max) | — | Alibaba | 0.00% |
| 26 | [Gemini 3.7 Flash](/models/gemini-3-7-flash) | high reasoning | Google | 0.00% |
| 27 | [Claude Opus 4.7](/models/claude-opus-4-7) | — | Anthropic | 0.00% |
| 28 | [GPT-5.6 Luna](/models/gpt-5-6-luna) | max reasoning | OpenAI | 0.00% |
| 29 | [DeepSeek V4 Flash 0731](/models/deepseek-v4-flash-0731) | high reasoning | DeepSeek | 0.00% |
| 30 | [GPT-5.4](/models/gpt-5-4) | xhigh reasoning | OpenAI | 0.00% |
| 31 | [GLM-5.3-Flash](/models/glm-5-3-flash) | max reasoning | Z.AI | 0.00% |
| 32 | [Grok 4.5](/models/grok-4-5) | high reasoning | xAI | 0.00% |
| 33 | [GPT-5.4 mini](/models/gpt-5-4-mini) | xhigh reasoning | OpenAI | 0.00% |
| 34 | [Gemini 3.6 Flash](/models/gemini-3-6-flash) | high reasoning | Google | 0.00% |
| 35 | [Muse Spark 1.1](/models/muse-spark-1-1) | xhigh reasoning | Meta | 0.00% |
| 36 | [Gemini 3.5 Flash](/models/gemini-3-5-flash) | high reasoning | Google | 0.00% |
| 37 | [GLM-5.1](/models/glm-5-1) | — | Z.AI | 0.00% |
| 38 | [Gemini 3.1 Pro Preview](https://www.vals.ai/models/google_gemini-3.1-pro-preview) | high reasoning | Google | 0.00% |
| 39 | [Qwen3.6 Plus](/models/qwen3-6-plus) | — | Alibaba | 0.00% |
| 40 | [Qwen3.8-27B](/models/qwen3-8-27b) | xhigh reasoning | Alibaba | 0.00% |
| 41 | [Kimi K2.7 Code](/models/kimi-k2-7-code) | — | Moonshot AI | 0.00% |
| 42 | [Kimi K2.6](/models/kimi-2-6) | — | Moonshot AI | 0.00% |
| 43 | [Nemotron 3 Ultra 550b A55b](https://www.vals.ai/models/nvidia_nemotron-3-ultra-550b-a55b) | — | Nvidia | 0.00% |
| 44 | [DeepSeek V4 Pro 0813](/models/deepseek-v4-pro-0813) | max reasoning | DeepSeek | 0.00% |
| 45 | [Gemini 3 Flash Preview](https://www.vals.ai/models/google_gemini-3-flash-preview) | high reasoning | Google | 0.00% |
| 46 | [Gemini 3.5 Flash-Lite](/models/gemini-3-5-flash-lite) | high reasoning | Google | 0.00% |
| 47 | [Claude Haiku 4.5 Thinking](/models/claude-haiku-4-5-thinking) | — | Anthropic | 0.00% |
| 48 | [Grok 4.3](/models/grok-4-3) | high reasoning | xAI | 0.00% |
| 49 | [Laguna XS.2](/models/laguna-xs-2) | — | Poolside | 0.00% |
| 50 | [Inkling](/models/inkling) | 0.99 reasoning | Thinking Machines Lab | 0.00% |
| 51 | [Laguna M.1](/models/laguna-m-1) | — | Poolside | 0.00% |
| 52 | [Gemini 3.1 Flash Lite Preview](https://www.vals.ai/models/google_gemini-3.1-flash-lite-preview) | high reasoning | Google | 0.00% |
| 53 | [MiniMax M2.7](/models/minimax-m2-7) | — | MiniMax | 0.00% |

## FAQ

### What does ProgramBench measure?

Can language models rebuild programs from scratch?

### Which model leads the published ProgramBench snapshot?

Claude Opus 5.5 currently leads the published ProgramBench snapshot with a score of 18.50%.

### How many models are evaluated on ProgramBench?

The September 27, 2026 contains 53 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.