ProgramDistill Full Reconstruction: leaderboard
Whole-application track of Microsoft Research's ProgramDistill (September 2026): starting from a minimal scaffold and a running reference, the agent rebuilds each of 12 web applications, then 413 cumulative workflows (1,541 replay stages) are replayed against the result. Workflow success is the percentage of those workflows that pass end to end; higher is better. Three models, each at max reasoning effort, averaging roughly $250 to $600 of inference per application.
Metric: Workflow success (%). Source: microsoft.github.io. Saturation forecast: Around January 2028. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (Max) | 49.15 | #6 (GPT-6 Astra) |
| 2 | Claude Opus 5 (Max) | 28.81 | #20 (Claude Opus 5) |
| 3 | GPT-5.6 Sol (Max) | 21.07 | #29 (GPT-5.6 Sol) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=programdistill-full-reconstruction · How It Works · Data refreshed daily, snapshot 2026-11-12.