The world around you is designed in CAD. How well can AI build it? Introducing CADArena, benchmarking AI CAD generation across the tools engineers use Agents can now build accurate geometry, but struggle to create feature trees engineers can maintain and edit
94
131
1,375
356,406
Normal retweeted
Today, @normalfactoryco joins the Specialized Intelligence Index with CAD Arena. Can AI turn an engineering drawing into a CAD part that's accurate and editable? Normal tests agents across five CAD platforms, expanding the SII into engineering design.
4
5
20
3,001
Opus 5.5 is the new #1 on CADArena. We added six new models. Opus 5.5 scores 0.750 and takes the top spot from GPT-6 Astra (0.671). And Grok 4.7 scores below Grok 4.6 and GPT-6 Luna.
The world around you is designed in CAD. How well can AI build it? Introducing CADArena, benchmarking AI CAD generation across the tools engineers use Agents can now build accurate geometry, but struggle to create feature trees engineers can maintain and edit
11
27
411
54,148
Each model gets an engineering drawing of a real part and rebuilds it natively in SolidWorks, NX, Onshape, Fusion and Build123d. We score the geometry and whether the feature tree is one an engineer could actually edit. Full results: normal.ai/leaderboard/cad-ar…
3
1,272
GPT-6 Astra costs 59.7% less than Fable 5.1 overall, with a slightly higher mean score: 0.671 versus 0.662
2
2
77
7,014
AI makes parts, but not in the way a human would Eg. Fable 5.1 fails to do a linear pattern, and then making 106 individual cuts instead
3
2
86
43,627
Mean FrontierCAD scores across platforms range from 0.414 to 0.449 across five platforms, while estimated cost per trial varies by more than 2×
1
2
62
8,570
The world around you is designed in CAD. How well can AI build it? Introducing CADArena, benchmarking AI CAD generation across the tools engineers use Agents can now build accurate geometry, but struggle to create feature trees engineers can maintain and edit
94
131
1,375
356,406
hellooooo world
18
83
15,894