Start with a model
that fits the work.

This is our recommendation for the desktop's reviewed agent workflow. Both models accept image context and support reasoning. The app still controls the proposed actions and asks for approval; model capability does not grant permission to operate your CAD software.

Standard USD prices per million text tokens, before caching or other adjustments. Screenshots, reasoning, repeated steps and retries can add usage. See OpenAI's Sol documentation and Astra documentation for current capabilities and pricing.

Your provider charges you directly. A ChatGPT subscription is separate from API usage. Set a provider spending limit before running tasks.

We compared the same reasoning setting rather than comparing a maximum-effort score with a medium-effort setup. Artificial Analysis reports Intelligence Index scores of 48 for Sol medium and 50 for Astra medium. The lower-cost Sol is our practical beta recommendation; Astra is our quality-first option.

Sources: Sol medium analysis, Astra medium analysis and the full model comparison.

That index measures general reasoning, knowledge, maths and coding. It is not a benchmark of your SOLIDWORKS, Logikal or Onshape task, and it does not certify a model for engineering decisions.

The desktop beta uses its own bounded screenshot-and-action workflow. Native provider computer-use tools are a separate integration; support in the model documentation does not mean every such tool is enabled in this app.

1. Add your own provider key in the desktop app.
2. Select the recommended model. The app uses medium reasoning automatically for Sol and Astra.
3. Use a copy of a model or drawing you can inspect.
4. Give a bounded brief, review the proposed plan and approve the actions.
5. Inspect the output before saving or issuing any engineering work.

You can choose a different supported model. Provider access, region and availability vary, and these recommendations will change as models and our own workflow checks improve.