SciDesignBench (20-Turn Feedback): leaderboard
Metric: Long-horizon de novo design, one attempt of up to 20 design-simulate-revise rounds with structured simulator feedback: mean success rate (%) over the 10 shared-core domains of SciDesignBench v1 (20 tasks per domain, 5 per difficulty level, temperature 0.7); a task succeeds when the model's JSON design, run through the domain's forward simulator, meets every target and constraint; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 76 | #60 |
| 2 | Claude Sonnet 4.6 | 71.5 | #85 |
| 3 | Claude Sonnet 4.5 | 69.5 | #138 |
| 4 | GPT-5.2 | 67 | #105 |
| 5 | Gemini 3.1 Pro (Preview) | 60.5 | #54 |
| 6 | Gemini 2.0 Flash | 60.5 | #331 |
| 7 | GPT-4o | 55 | #333 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=scidesignbench-20-turn-feedback · How It Works · Data refreshed daily, snapshot 2026-10-11.