ProgramDistill
Agents inspect working web applications and restore missing features in editable copies. Replayed browser traces check whether patches recover the reference behavior. The paper evaluates partial repairs and full application reconstruction.
The paper constructs 4,063 tasks across 26 applications and uses a 300-task subset for partial reconstruction. It states that public release is in progress. Restoration depth and lineage scoring define distinct reporting settings.
- Made by
- Microsoft Research, Microsoft AI, and KAIST
- Released
- September 2026
- Kind
- Web and app front ends
- Level
- University
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- Kept private
- Measures
- Capability
- Language
- English
- Interaction
- Agentic
- Items from
- Real work and use
- Refresh
- One fixed set
- Run by
- A host
- In views
- Computer use, Agents
- Strongest model tested
- GPT-6 Astra #4 on the capability index
- Last model added
- September 2026
- Scores refreshed
- November 20th, 2026
In the authors’ words
Coding agents are typically evaluated with desired behavior specified through issues or instructions. In practical web development, however, agents may need to infer behavior from working software and implement it in an incomplete application. We introduce ProgramDistill, a benchmark evaluating coding agents on features discovered through interaction with fully functional reference applications. We build ProgramDistill by factorizing applications into features of different granularities, each associated with replayable behaviors executable via its gold patch. Our pipeline, mine-craft-patch, discovers 1,975 replay-verified behaviors across 26 applications and constructs 4,063 tasks without human intervention. Across nine frontier coding agents, GPT-6 Astra and Claude Opus 5 achieve 49.2% and 28.8% success on cumulative workflows in full-application reconstruction. In partial-application reconstruction, success falls from 100% to 64.0% and from 96% to 32% as restoration depth increases from 1 to 8. ProgramDistill thus provides a scalable benchmark with controlled difficulty for evaluating and diagnosing coding agents, and a natural basis for future curriculum-based training.
GPT-6 Astra leads with 84.3% among models released by September 2026
Best score per model from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best.
The newest model with a score here came out in September 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- Claude Opus 5.5
- and 13 more
- 1 GPT-6 Astra ProgramDistill-300; partial reconstruction; high effort; 1000 steps 84.3%
- 2 Claude Opus 5 ProgramDistill-300; partial reconstruction; high effort; 1000 steps 68.7%
- 3 GPT-5.6 Sol ProgramDistill-300; partial reconstruction; high effort; 1000 steps 60.7%
- 4 Grok 4.6 ProgramDistill-300; partial reconstruction; high effort; 1000 steps 48.3%
- 5 Claude Sonnet 5 ProgramDistill-300; partial reconstruction; high effort; 1000 steps 47.3%
- 6 GPT-5.3 Codex ProgramDistill-300; partial reconstruction; high effort; 1000 steps 45.7%
- 7
Gemini 3.7 Flash ProgramDistill-300; partial reconstruction; high effort; 1000 steps 45.3%
- 8
Gemini 3.6 Flash ProgramDistill-300; partial reconstruction; high effort; 1000 steps 29.7%
- 9
Gemini 3.1 Pro ProgramDistill-300; partial reconstruction; high effort; 1000 steps 22.3%
Source: The ProgramDistill paper
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-6 Astra | 84.3% | ProgramDistill-300; partial reconstruction; high effort; 1000 steps | Sep 2026 | Official leaderboard |
| 2 | Claude Opus 5 | 68.7% | ProgramDistill-300; partial reconstruction; high effort; 1000 steps | Jul 2026 | Official leaderboard |
| 3 | GPT-5.6 Sol | 60.7% | ProgramDistill-300; partial reconstruction; high effort; 1000 steps | Jul 2026 | Official leaderboard |
| 4 | Grok 4.6 | 48.3% | ProgramDistill-300; partial reconstruction; high effort; 1000 steps | Aug 2026 | Official leaderboard |
| 5 | Claude Sonnet 5 | 47.3% | ProgramDistill-300; partial reconstruction; high effort; 1000 steps | Jun 2026 | Official leaderboard |
| 6 | GPT-5.3 Codex | 45.7% | ProgramDistill-300; partial reconstruction; high effort; 1000 steps | Feb 2026 | Official leaderboard |
| 7 | | 45.3% | ProgramDistill-300; partial reconstruction; high effort; 1000 steps | Aug 2026 | Official leaderboard |
| 8 | | 29.7% | ProgramDistill-300; partial reconstruction; high effort; 1000 steps | Jul 2026 | Official leaderboard |
| 9 | | 22.3% | ProgramDistill-300; partial reconstruction; high effort; 1000 steps | Feb 2026 | Official leaderboard |
Rankings by part
We read 3 parts from the benchmark’s paper: Cumulative tasks, Logic only and Logic and UI. Each ranks models on one part of the test. GPT-6 Astra leads every one of them. Arrows show how far a model moved from its overall rank among the models we track.
Cumulative tasks
9 models we track have a score here.
- 1 GPT-6 Astra 81.2%
- 2 Claude Opus 5 63.2%
- 3 GPT-5.6 Sol 54.4%
- 4 Claude Sonnet 5 ↑1 40.8%
- 5 Grok 4.6 ↓1 40.4%
- 6 GPT-5.3 Codex 38.8%
- 7
Gemini 3.7 Flash 37.6%
- 8
Gemini 3.6 Flash 23.2%
- 9
Gemini 3.1 Pro 12.8%
Logic only
9 models we track have a score here.
- 1 GPT-6 Astra 92.9%
- 2 Claude Opus 5 80%
- 3 GPT-5.6 Sol 76.4%
- 4 Claude Sonnet 5 ↑1 65%
- 5 Grok 4.6 ↓1 64.3%
- 6
Gemini 3.7 Flash ↑1 61.4%
- 7 GPT-5.3 Codex ↓1 59.6%
- 8
Gemini 3.6 Flash 37.9%
- 9
Gemini 3.1 Pro 31.1%
Logic and UI
9 models we track have a score here.
- 1 GPT-6 Astra 76.9%
- 2 Claude Opus 5 58.8%
- 3 GPT-5.6 Sol 46.9%
- 4 Grok 4.6 34.4%
- 5 GPT-5.3 Codex ↑1 33.4%
- 6 Claude Sonnet 5 ↓1 31.9%
- 7
Gemini 3.7 Flash 31.2%
- 8
Gemini 3.6 Flash 22.5%
- 9
Gemini 3.1 Pro 14.7%
2 models have held number one in the benchmark’s paper since its release in September 2026
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The ProgramDistill paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official ProgramDistill leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Coding
- SWE-bench Verified In use 500 human-validated GitHub issues from 12 Python repos. Agent writes a patch that must pass hidden tests. Claude Opus 4.7 83.5%
- LiveCodeBench v6 In use Competitive programming problems collected after model cutoffs from LeetCode, AtCoder, Codeforces. Scored pass@1 on hidden tests.
DeepSeek-V4-Pro 93.5%
- SWE-bench Multilingual In use 300 GitHub-issue tasks across nine programming languages in SWE-bench format. Patch graded by repository tests. Claude Fable 5.2 96.6%
- DeepSWE In use 113 original long-horizon engineering tasks across 91 repos and five languages, written from scratch. Hand-written verifiers accept alternative solutions. GPT-6 Astra 74.1%
- SciCode In use Scientist-curated research coding problems split into subproblems across natural sciences. Code checked against test cases.
Seed2.1 Pro 59.8%
- Codeforces rating In use Model solves recent Codeforces contest problems. An Elo-style rating is computed from simulated contest performance.
DeepSeek V4.1 Flash 3,471
Data audit
All 10 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access The tasks are private.
- Pass Confidence Checked against a primary source.
Scores 4 pass 1 note
- Pass Independent runs Scores from the benchmark’s paper (9 entries).
- Pass Models scored 9 models we track.
- Pass Newest score The newest is from September 2026, for GPT-6 Astra.
- Note Frontier models Only 1 of the top 10 has a score. The strongest tested is GPT-6 Astra (#4).
- Pass Headroom The best is 84.3%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 9 rows for models we track at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.