Vals AI ProgramBench: leaderboard
Each of the 200 tasks hands the model an execute-only command-line binary and a behaviour specification and asks for a cleanroom reimplementation, with no internet, source lookups or decompilers. Vals AI runs them in a mini-SWE-agent harness with a bash tool in an offline sandbox, then compiles each submission and runs hidden tests. This board is the average share of those tests passed per task, higher is better.
Metric: Raw Pass Rate (%). Source: www.vals.ai. Saturation forecast: Around August 2028. 64 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 | 18.5 | #3 |
| 2 | Claude Fable 5.1 | 7 | #9 |
| 3 | Claude Sonnet 5.5 | 6.5 | #11 |
| 4 | GPT-6 Astra | 5.5 | #6 |
| 5 | Claude Opus 5 | 3 | #12 |
| 6 | GPT-6.1 Sol | 3 | #7 |
| 7 | Gemini 4 Argon | 2.5 | #8 |
| 8 | Claude Haiku 5.5 | 2.5 | #40 |
| 9 | Muse Spark 1.3 (Max) | 2.5 | #24 (Muse Spark 1.3) |
| 10 | Kimi K3 | 2 | #30 |
| 11 | Claude Fable 5 | 2 | #16 |
| 12 | GPT-6 Sol | 2 | #14 |
| 13 | GPT-5.6 Sol | 1.5 | #19 |
| 14 | GLM-5.3 | 1.5 | #39 |
| 15 | Claude Opus 4.8 | 1 | #37 |
Interactive version: theaggregate.ai/benchmark?slug=vals-ai-programbench · How It Works · Data refreshed daily, snapshot 2026-10-20.