CADEngBench - Physics Reach (L3): leaderboard

Metric: FEA reach rate (%; share of the 164 L3-eligible generated parts whose geometry reaches a completed CalculiX linear-static solve under the benchmark's fixed material, supports and loads; reaching a solve does not require matching the reference response). Source: arxiv.org. Saturation forecast: Around April 2027. 8 models tracked.

Top models

#ModelScore
1Claude Sonnet 4.554.3
2Kimi K2.5 (Non-reasoning)51.8
3Llama 4 Maverick47.6
4Gemini 3 Flash (Preview) (Minimal)46.3
5GLM-4.6V (Non-reasoning)44.5
6GPT-5.2 (Non-reasoning)37.8
7Qwen 3.5 35B A3B (Non-reasoning)15.9

Interactive version: theaggregate.ai/benchmark?slug=cadengbench-physics-reach-l3 · How It Works · Data refreshed daily, snapshot 2026-09-29.