VASPBench (AutoDFT-Full): leaderboard
VASPBench (Yang et al., 2026): 34 density functional theory calculation tasks taken from the official VASP documentation, where a crystal structure and a plain-language objective must be turned into converged VASP outputs that pass the benchmark's validation checks. The LLM drives the closed-loop AutoDFT-Full agent, which plans each step, sets parameters from earlier results, monitors running jobs, repairs failures and revises the plan; score is the percentage of tasks solved in one run, higher is better.
Metric: Task success rate (%; converged outputs passing the validation checks on 34 DFT tasks, closed-loop AutoDFT-Full, one run per task). Source: arxiv.org. Saturation forecast: Estimated already saturated. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 94.1 | #172 |
| 2 | Gemini 3.1 Pro (Preview) | 91.2 | #91 |
| 3 | Claude Sonnet 4.6 | 85.3 | #158 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=vaspbench-autodft-full · How It Works · Data refreshed daily, snapshot 2026-11-12.