URSA Single-Step Retrosynthesis: leaderboard
Metric: Average per-target ChemCensor score (0 to 5) over the model's top 10 unique predicted reactant sets (exact duplicates removed, missing slots scored 0), averaged over the 100 expert-validated novel drug-like targets of URSA-expert-2026 (15 independent samples per target, five-shot prompts drawn from USPTO-full training reactions, no tools, reasoning enabled at the default budget for models that support it; ChemCensor 0.5.2 scores each reaction from 0 (no precedent) to 5 by the substructural context of its reaction centre and functional groups found in USPTO-full precedents); higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4.1 (Thinking) | 0.49 | #218 (Grok 4.1) |
| 2 | Gemini 3 Flash (Preview) | 0.42 | #78 |
| 3 | Claude Sonnet 4.5 (Thinking) | 0.29 | #138 (Claude Sonnet 4.5) |
| 4 | Claude Opus 4.5 (Thinking) | 0.23 | #79 (Claude Opus 4.5) |
| 5 | Kimi K2 | 0.19 | #236 |
| 6 | GPT-5.2 (Thinking) | 0.15 | #105 (GPT-5.2) |
| 7 | GPT-5.1 (Thinking) | 0.11 | #131 (GPT-5.1) |
| 8 | Gemini 2.5 Flash | 0.09 | #237 |
| 9 | DeepSeek V3.2 (Thinking) | 0.05 | #198 (DeepSeek V3.2) |
| 10 | Qwen 3 8B | 0 | #667 |
| 11 | Qwen 3 14B | 0 | #524 |
| 12 | LFM2 2.6B | 0 | #1286 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=ursa-single-step-retrosynthesis · How It Works · Data refreshed daily, snapshot 2026-10-11.