ScienceBoard-Verified - Lean 4: leaderboard
ScienceBoard-Verified scientific computer-use tasks across KAlgebra, Celestia, GRASS GIS, ChimeraX, Lean 4 and TeXstudio; source success rate, with Average the unweighted mean of six applications. Model-based API harness uses the latest three steps plus the current screenshot/accessibility tree; native Grok Build/Codex maintain persistent per-task sessions, with web search disabled. Resource construction and harness identities remain separate run configurations. Lean 4.
Metric: Success rate (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 46 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen-3.8 Max (xhigh) (ScienceCUA Run: harness=model-based; resource=skillfoundry) | 34.62 | |
| 2 | GPT-5.5 (xhigh) (ScienceCUA Run: harness=model-based; resource=Full%20ScienceCUA; evolution=108%20additional%20tasks%3B%20frozen%20library) | 34.62 | |
| 3 | GPT-5.5 (xhigh) (ScienceCUA Run: harness=model-based; resource=SkillCAT; evolution=108%20additional%20tasks%3B%20frozen%20library) | 34.62 | |
| 4 | Qwen-3.8 Max (xhigh) (ScienceCUA Run: harness=model-based; resource=Vanilla) | 30.77 | |
| 5 | Qwen-3.8 Max (xhigh) (ScienceCUA Run: harness=model-based; resource=RAG) | 30.77 | |
| 6 | Qwen-3.8 Max (xhigh) (ScienceCUA Run: harness=model-based; resource=visualskill) | 30.77 | |
| 7 | Qwen-3.8 Max (xhigh) (ScienceCUA Run: harness=model-based; resource=ScienceCUA) | 30.77 | |
| 8 | GPT-5.5 (xhigh) (ScienceCUA Run: harness=model-based; resource=RAG) | 30.77 | |
| 9 | Grok-4.6 (xhigh) (ScienceCUA Run: harness=Grok%20Build; resource=skillfoundry) | 30.77 | |
| 10 | GPT-5.5 (xhigh) (ScienceCUA Run: harness=Codex; resource=Vanilla) | 30.77 |
Interactive version: theaggregate.ai/benchmark?slug=scienceboard-verified-lean-4 · How It Works · Data refreshed daily, snapshot 2026-11-05.