MolDesignBench - Molecule Optimization (Without Tools): leaderboard
The optimization task of MolDesignBench (Jeong et al., 2026): the agent must modify a given starting molecule so that it satisfies every implicit and explicit constraint while staying structurally similar, or answer None when the request is infeasible, without tools and relying only on its own reasoning. Success rate in percent over 1,000 instances, one tenth of them infeasible; higher is better.
Metric: Success rate (%; share of the 1,000 optimization instances handled correctly: a valid SMILES satisfying every implicit and explicit property and functional-group constraint on the 900 feasible instances, staying similar to the given starting molecule, or an explicit None on the 100 infeasible ones; agent without tools, from the model's own reasoning, at most 20 turns). Source: arxiv.org. Saturation forecast: Around 2029. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 17 | #121 |
| 2 | Claude Opus 4.6 (Thinking) | 17 | #100 (Claude Opus 4.6) |
| 3 | Qwen 3 235B A22B (Thinking) | 14 | #622 (Qwen 3 235B A22B) |
| 4 | Intern-S1 | 13 | #616 |
| 5 | GPT-OSS-120B | 11 | #699 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=moldesignbench-molecule-optimization-without-tools · How It Works · Data refreshed daily, snapshot 2026-11-12.