BenchCAD 2.1 (Anthropic runs, Claude Code): leaderboard

Models write programmatic CAD for the 100 held-out BenchCAD 2.1 Core cases across six tasks (parts and assemblies from engineering drawings or rendered views, board wiring from renders); the score (0 to 1) is the mean over passes when Anthropic ran each Claude model inside Claude Code. Unlike the BenchCAD figures in OpenAI's and Kimi's launch posts, which the BenchCAD site files under its 1.0 Vision2Code voxel-IoU task, this is the 2.1 Core set, with Anthropic's runs of four Claude models at five efforts.

Metric: Core score (0 to 1) on the 100 held-out BenchCAD 2.1 Core cases (six programmatic-CAD tasks), run in Claude Code, mean of passes; Anthropic's runs. Source: benchcad.com. Saturation forecast: Around May 2027. 20 models tracked.

Top models

#ModelScoreOverall rank
1Claude Opus 5.5 (Max)0.65#5 (Claude Opus 5.5)
2Claude Sonnet 5.5 (Max)0.61#19 (Claude Sonnet 5.5)
3Claude Opus 5.5 (xHigh)0.55#5 (Claude Opus 5.5)
4Claude Fable 5.1 (Max)0.44#15 (Claude Fable 5.1)
5Claude Fable 5.1 (xHigh)0.41#15 (Claude Fable 5.1)
6Claude Sonnet 5.5 (xHigh)0.39#19 (Claude Sonnet 5.5)
7Claude Opus 5.5 (High)0.3#5 (Claude Opus 5.5)
8Claude Fable 5.1 (High)0.26#15 (Claude Fable 5.1)
9Claude Haiku 5.5 (Max)0.23#65 (Claude Haiku 5.5)
10Claude Opus 5.5 (Medium)0.19#5 (Claude Opus 5.5)
11Claude Haiku 5.5 (xHigh)0.18#65 (Claude Haiku 5.5)
12Claude Sonnet 5.5 (High)0.17#19 (Claude Sonnet 5.5)
13Claude Fable 5.1 (Medium)0.16#15 (Claude Fable 5.1)
14Claude Fable 5.1 (Low)0.12#15 (Claude Fable 5.1)
15Claude Haiku 5.5 (High)0.11#65 (Claude Haiku 5.5)

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #6 GPT-6 Astra.

Interactive version: theaggregate.ai/benchmark?slug=benchcad-2-1-anthropic-runs-claude-code · How It Works · Data refreshed daily, snapshot 2026-11-13.