CAD-Bench
← back
PARA-005 · Paraphrase Robustness · difficulty 4/5

5× paraphrased planetary carrier

sha256:2b97cc4d1ef0aa55…

§1Prompt verbatim

Disc Ø 80 × 8 mm with central Ø 12 H7 bore, three Ø 6 H7 satellite bores on a 30 mm PCD at 0/120/240°, six M3 tapped holes on a 60 mm PCD at 30° offset.

§2Ground-truth spec

shells1
watertighttrue
manifoldtrue
acceptance ε±0.05 mm
Reference geometry is identical to MECH-027; this task evaluates paraphrase variance, not modelling skill.

§3Reference render

canonical reference · drag to orbit, scroll to zoom

Visualisation is rebuilt in-browser from the canonical parametric description. Scoring is performed against the held-out reference STEP file (sha-256 fingerprint above).

§4Per-agent renders

reference + 10 agent outputs · scored against the held-out STEP
vol IoU · BREP · manifold
canonical reference
REFERENCE
canonical · ground truth
1.000100✓
OpenAI o4 (reasoning) → CadQuery
OpenAI o4 (reasoning) → CadQuery
OpenAI + CadQuery 2.4
0.78611✓
Human Baseline (Mech-E)
Human Baseline (Mech-E)
n=4 senior engineers
0.72110✓
Claude Opus 4.7 → OpenSCAD
Claude Opus 4.7 → OpenSCAD
Anthropic + OpenSCAD 2024.06
0.7030✗
Adam (CADcrush)
Adam (CADcrush)
CADcrush
0.65010✗
Claude Opus 4.7 → CadQuery
Claude Opus 4.7 → CadQuery
Anthropic + CadQuery 2.4
0.64710✗
Zoo Text-to-CAD
Zoo Text-to-CAD
Zoo (KittyCAD)
0.61711✓
GPT-5 → CadQuery
GPT-5 → CadQuery
OpenAI + CadQuery 2.4
0.6179✗
Claude Sonnet 4.6 → CadQuery
Claude Sonnet 4.6 → CadQuery
Anthropic + CadQuery 2.4
0.59011✗
Trellis 3D
Trellis 3D
Microsoft Research
0.5690✗
Qwen3 Coder → CadQuery
Qwen3 Coder → CadQuery
Alibaba + CadQuery 2.4
0.53713✗
Gemini 2.5 Flash → CadQuery
Gemini 2.5 Flash → CadQuery
Google + CadQuery 2.4
0.51213✗
GPT-5 Mini → OpenSCAD
GPT-5 Mini → OpenSCAD
OpenAI + OpenSCAD 2024.06
0.48711✗
Claude Haiku 4.5 → CadQuery
Claude Haiku 4.5 → CadQuery
Anthropic + CadQuery 2.4
0.48513✗
DeepSeek R1 (reasoning) → CadQuery
DeepSeek R1 (reasoning) → CadQuery
DeepSeek + CadQuery 2.4
0.47714✗
Llama 3.3 70B → OpenSCAD
Llama 3.3 70B → OpenSCAD
Meta + OpenSCAD 2024.06
0.45315✗
Hunyuan3D-2
Hunyuan3D-2
Tencent
0.44311✗
Spline AI
Spline AI
Spline.design
0.4380✗
DeepCAD
DeepCAD
Wu et al. 2021 (research)
0.27919✗
no manifold solid produced
Gemini 2.5 Pro → OpenSCAD
Gemini 2.5 Pro → OpenSCAD
Google + OpenSCAD 2024.06
—18✗
no manifold solid produced
CAD-Coder R1
CAD-Coder R1
CAD-Coder Labs (research)
—74✗

Each tile is rebuilt from the canonical parametric description and degraded to match the agent's scored profile (tessellation, non-manifold face removal, dimension scale jitter, missing features). Image-only diffusion models render visually plausible meshes but score in the single digits on BREP fidelity — the geometry is not a manifold solid even when the render reads clean.

§5Per-agent metrics

ranked by Vol IoU · same data as the leaderboard, restricted to this task
AgentWatert.Manif.Paraphrase IoU σSeed σP@1p50latencycost
OpenAI o4 (reasoning) → CadQuery✓0.9610.0200.0361.000—115.0s$1.214
Human Baseline (Mech-E)✓0.9620.0070.0060.000—733.0s$6.526
Claude Opus 4.7 → OpenSCAD✓0.9490.0360.0340.000—23.4s$0.337
Adam (CADcrush)✓0.9470.0560.0260.000—10.5s$0.229
Claude Opus 4.7 → CadQuery✓0.9420.0300.0490.000—47.8s$0.362
Zoo Text-to-CAD✓0.9500.0270.0240.000—7.2s$0.149
GPT-5 → CadQuery✓0.9380.0430.0430.000—48.6s$0.175
Claude Sonnet 4.6 → CadQuery✓0.9460.0300.0570.000—20.5s$0.060
Trellis 3D✓0.9350.0510.1220.000—12.8s$0.052
Qwen3 Coder → CadQuery✓0.9320.0590.0540.000—17.5s$0.031
Gemini 2.5 Flash → CadQuery✓0.9250.0420.0720.000—11.6s$0.017
GPT-5 Mini → OpenSCAD✓0.9270.0420.0680.000—17.2s$0.012
Claude Haiku 4.5 → CadQuery✓0.9250.0520.0420.000—8.0s$0.019
DeepSeek R1 (reasoning) → CadQuery✓0.9250.0400.0690.000—74.2s$0.038
Llama 3.3 70B → OpenSCAD✓0.9220.0770.0450.000—27.5s$0.024
Hunyuan3D-2✓0.9220.0500.0740.000—29.5s$0.059
Spline AI✓0.9180.0680.0730.000—8.1s$0.047
DeepCAD×0.8950.1020.0590.000—5.1s$0.022
Gemini 2.5 Pro → OpenSCAD
kernel error: BRepCheck_NotClosed
×0.000——0.000—33.0s$0.104
CAD-Coder R1
kernel error: BRepCheck_NotClosed
×0.000——0.000—6.7s$0.005