RESULTS · OCTOBER 2026
Argus CAD Benchmark
We handed Argus CAD tasks engineers usually do manually, from modeling parts to reviewing drawings, and let it work on its own.
How to read this page
What am I looking at?
Results from a benchmark we ran on Argus, Cosmon’s AI agent, running on Gemini 4 Argon. Here’s how to read the page.
- Each card is one real CAD task that Argus did on its own, in the CAD tool named on the card. Its front shows level of autonomy (the share of the task’s checks met with no one stepping in), time on task (minutes from starting work to handing back a result) and cost (model usage, as tracked on our platform).
- Time on task charts Argus’s minutes on every task. Manual effort per task sets them against the hours an engineer would need manually (estimated) and how demanding each task is.
- Click any bar or row to open that task’s card. Flip a card for its checks and a with-Argus vs fully manual comparison; Benchmark details has the task’s objective and every check with its target and result.
- How the benchmark was run, and which numbers are measured and which are estimated, is under Method and limits at the bottom of the page.
Time on task
15MINmedian time on task, across all 21 tasks
No task had a step limit.
Manual effort per task
96Hof manual work across all 21 tasks (est.).
Argus took 6.5 h in total.
Argus took 6.5 h in total.
Manual time is estimated. Cognitive load (0–100) is rated by an AI judge using NASA-TLX.
What Argus delivered
Flip a card to see the results
Swipe for more tasks
Pictures show each task’s model, from the benchmark’s own reference files. Line drawings are illustrative. How the benchmark was run is under Method and limits, below.
How we ran it
Method and limits
Setup
- Agent
- Argus on Gemini 4 Argon, at its standard reasoning setting
- Tasks
- 21 tasks: 12 in SolidWorks, 8 in NX and 1 in Creo. 18 are the same tasks as on the Opus 5.5 page.
- Limit
- No step limit, and no one stepped in on any task. A step is one action in the CAD software.
- Grading
- Expert answer keys, checked automatically. Where a task has no checker, an AI model grades each check against the answer key. An AI judge scores drawing reviews.
- Source
- One scorecard records every task’s time, steps and checks. Every number on this page comes from it.
Measured
- Time on taskStart of work to the handed-back result
- Level of autonomyShare of checks met with no one stepping in
- CostModel usage, as tracked on our platform
Estimated
- Manual timeFor a mid-level mechanical engineer
- Engineer reviewPer task, rated by an AI judge from what has to be checked
- Cognitive loadRated by an AI judge using NASA-TLX
