PDEAgent-Bench - Firedrake: leaderboard
PDEAgent-Bench (Hang et al., 2026) gives a model a structured PDE specification and asks for a complete finite-element solver in one attempt, with no execution feedback, here for the Python library Firedrake: the 510 cases of eight PDE families that the Firedrake track supports. A case passes only if the solver runs, returns a valid solution within 10 times the calibration baseline's relative L2 error and finishes within 3 times its runtime; the score is the pass rate in percent, higher is better.
Metric: Case pass rate (%; 510 Firedrake cases, single-shot, temperature 0; same pass gates). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 45.9 | #75 |
| 2 | GPT-5.4 | 38 | #121 |
| 3 | Gemini 3.1 Pro (Preview) | 32.4 | #91 |
| 4 | Qwen 3.6 Plus | 22.9 | #217 |
| 5 | DeepSeek V3.2 | 2.7 | #403 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=pdeagent-bench-firedrake · How It Works · Data refreshed daily, snapshot 2026-11-12.