[KongchangAI]
Benchmark

CAD World

The first computer-use benchmark targeting long-horizon mechanical CAD workflows, comprising 200 tasks across 11 workflow categories. It evaluates computer-use agents on automated operation tasks within professional CAD software.

Timeline (last 90 days)

Sep 26

CAD World is the first computer operation benchmark targeting long-horizon mechanical CAD workflows.

Unverified50%
Sep 26

In CAD World's 200-task evaluation, the best-performing GPT-5.4 achieved a success rate of only 17.5%.

Unverified50%
Sep 26

In the CAD World benchmark, human experts achieved a success rate of 87%, requiring an average of only 26.5 steps.

Unverified50%
Sep 26

CAD World tested seven mainstream computer operation agents.

Unverified50%
Sep 26

GPT-5.4 achieved a completion rate of 62.5% on CAD World, with an average of 104.6 action steps and an average of 46.9 steps for successfully completed tasks.

Unverified50%
Sep 26

Opus-4.8 achieved a success rate of 16% on CAD World, with an average cost of $10 per task.

Unverified50%
Sep 26

Kimi K2.6 achieved a success rate of 7.5% on CAD World, OpenCUA achieved 1.5%, and two other models achieved a success rate of 0.

Unverified50%
Sep 26

CAD World contains 200 tasks, covering 11 workflow categories and 183 distinct knowledge points.

Unverified50%
Sep 26

In CAD World, the sketch category contains 63 tasks and the part modeling category contains 77 tasks.

Unverified50%
Sep 26

CAD World uses the open-source FreeCAD as its runtime environment, with FreeCAD running inside a pre-built Ubuntu virtual machine managed via Docker and QEMU.

Unverified50%

3 more timeline events

All Facts (13)

Unverified

CAD World is the first computer operation benchmark targeting long-horizon mechanical CAD workflows.

50%
Unverified

In CAD World's 200-task evaluation, the best-performing GPT-5.4 achieved a success rate of only 17.5%.

50%
Unverified

In the CAD World benchmark, human experts achieved a success rate of 87%, requiring an average of only 26.5 steps.

50%
Unverified

CAD World tested seven mainstream computer operation agents.

50%
Unverified

GPT-5.4 achieved a completion rate of 62.5% on CAD World, with an average of 104.6 action steps and an average of 46.9 steps for successfully completed tasks.

50%
Unverified

Opus-4.8 achieved a success rate of 16% on CAD World, with an average cost of $10 per task.

50%
Unverified

Kimi K2.6 achieved a success rate of 7.5% on CAD World, OpenCUA achieved 1.5%, and two other models achieved a success rate of 0.

50%
Unverified

CAD World contains 200 tasks, covering 11 workflow categories and 183 distinct knowledge points.

50%
Unverified

In CAD World, the sketch category contains 63 tasks and the part modeling category contains 77 tasks.

50%
Unverified

CAD World uses the open-source FreeCAD as its runtime environment, with FreeCAD running inside a pre-built Ubuntu virtual machine managed via Docker and QEMU.

50%
Unverified

The annotation and verification process for CAD World took approximately three months and 1,300 person-hours, with a median instruction length of 63 words.

50%
Unverified

In CAD World, each task is limited to a maximum of 100 steps, and agents cannot directly access files or call APIs.

50%
Unverified

As model capabilities improve, the failure stage in CAD tasks shifts later rather than disappearing entirely, with stronger models' failures attributable to document structure errors, geometric errors, and modeling workflow errors.

50%

Source Articles