All research
BenchmarkPreliminary
HappyCAD on BenchCAD 2.0 Public
HappyCAD scored 0.856 / 1 on the 91 CAD cases of the BenchCAD 2.0 Public set.

Results
| Task | HappyCAD (Public) | GPT-6 Astra | Opus 5.5 | Sonnet 5.5 | Fable 5.1 | Haiku 5.5 |
|---|---|---|---|---|---|---|
| 1. Drawing to CAD: part | 0.940 (21) | 0.807 | 0.837 | 0.793 | 0.521 | 0.561 |
| 2. Drawing to CAD: assembly, parts supplied | 0.971 (14) | 0.721 | 0.667 | 0.342 | 0.398 | 0.091 |
| 3. 3D reverse engineering: part | 0.955 (23) | 0.576 | 0.592 | 0.591 | 0.310 | 0.316 |
| 4. 3D reverse engineering: assembly | 0.626 (27) | 0.501 | 0.440 | 0.438 | 0.246 | 0.154 |
| 5. Drawing to CAD: parts and assembly | 0.947 (6) | 0.690 | 0.661 | 0.223 | 0.372 | 0.010 |
| All five tasks | 0.856 (91) | 0.612 | 0.589 | 0.455 | 0.330 | 0.196 |
- Scores: from 0 to 1, where 1.0 is a perfect answer. Each score is the mean over the task's cases.
- Scoring: BenchCAD's own scorer compares each answer with a reference model: the shape of each part, and the position of each part in an assembly.
- Cases: HappyCAD ran on the Public set (91 cases; the count for each task is in brackets). The five models in the table ran on the held-out Core set (88 cases: 6, 11, 28, 25 and 18 for tasks 1–5). The two sets have no cases in common.
- Model columns: BenchCAD's own results for BenchCAD 2.0, from its 25 published runs, each model at max effort.
- Other models: newer models, such as Claude Fable 5.2 and GPT-6.2 Astra, have only an overall BenchCAD 2.1 score (leaderboard), with no score for each task, so the table does not include them.
- Official score: only Core gives an official BenchCAD score. Our request for access to Core is in review with the BenchCAD team. When they approve it, we will run HappyCAD on Core.