GPT-6 Astra for CAD: run OpenAI's best CAD model in SolidWorks, Onshape, and Fusion

GPT-6 Astra is OpenAI's frontier model, released to approved organizations on September 3, 2026 and to ChatGPT Plus, Pro, Business, and Enterprise users the next day. It's also available through the OpenAI API, Microsoft Azure, and AWS Bedrock. OpenAI describes it as its most capable model for computer use, coding, scientific reasoning, and long-context work, and the launch post makes a point of 3D and CAD-style tasks: modeling a house in Blender and turning it into a walkable Unreal Engine 5 scene, and laying out a manufacturable PCB in KiCad from a schematic.

The reason engineers are searching for "GPT-6 Astra CAD" is the benchmarks. On BenchCAD, which asks a model to reconstruct a CAD program from rendered views, Astra reaches 95.9% geometric overlap when given tools, against 84.3% for Claude Fable 5.1 and 83.3% for GPT-5.6 Sol. On CAD Arena, an independent test that had six models build real components in Siemens NX, SolidWorks, Onshape, Autodesk Fusion, and Build123d, Astra averaged 0.671 to Fable 5.1's 0.662, with its best result in Onshape at 0.723. That's the first time an OpenAI model has led a native-CAD comparison, and it did so at roughly a third of the per-run cost of the runner-up.

Strip away the launch-week hype and the capability is real but specific. Astra is very good at turning a description or a reference image into geometry, and at driving CAD software through an API or an agent loop when it can render, measure, and iterate. Published demos and hands-on reports cover:

• Reconstructing parts and assemblies from multi-view renders and reference photos, the task BenchCAD measures
• Building multi-part assemblies in Onshape, including a turbofan with a visible assembly structure
• Taking a raw scanner mesh, remodeling it as a solid, and exporting a STEP file that opens cleanly in SOLIDWORKS
• Converting a 2D drawing of a steam locomotive into thousands of editable objects
• Driving Fusion through the Fusion MCP server from Codex, and Onshape through FeatureScript
• Laying out PCBs in KiCad and building scenes in Blender, per OpenAI's own launch examples

The pattern across all of these is the same: Astra does its best work when it sits on top of a real CAD kernel with a feedback loop. The BenchCAD score that made headlines is the with-tools number. A model that can sketch, extrude, render, compare, and try again behaves like an engineer iterating. A model answering in one shot does not. The harness around the model, the tools it can call and the checks it runs on its own output, is what turns a benchmark score into a part you can use.

Two caveats matter if you're about to trust it with production parts. First, geometric similarity is not design intent. One review describes a run that scored 96.1% geometric similarity while the requested modification had not actually been made. On editing benchmarks that resemble everyday design work, professional designers still beat frontier models by more than fifty percentage points.

Second, CAD Arena's own conclusion was that strong models "hit geometry well across platforms" but "fall back on editability." A part that looks right as a STEP export but arrives as an unconstrained blob in the feature tree is a part someone has to rebuild. The models that will change mechanical design are the ones that learn to produce a clean, parametric construction history, and that depends as much on the tools around the model as on the model itself.

Adam is a mechanical-engineering harness that runs on the model you choose. GPT-6 Astra sits in Adam's Ultra model family alongside the other frontier models, and every Ultra model gets the same CAD skill modules, the same native integrations with SolidWorks, Onshape, Fusion, and Rhino, and the same render-measure-check loop. Getting started takes a few minutes:

• Sign in at adam.new. Pro and Max plans unlock the Ultra family; Free includes Lite only.
• Connect your CAD. Onshape connects in the browser. SolidWorks, Fusion, and Rhino use Adam's desktop connector.
• In the composer, open Advanced, then Model, and pick GPT 6 Astra under Ultra. Or leave the preset on Ultra and let Adam choose.
• Describe the part, edit, or assembly. Adam builds it in your CAD document with a real feature tree, checks the result, and iterates.
• Review the change in place. Adam works in your CAD file, so there's nothing to export and re-import.

Prefer to stay in Claude, Codex, or Cursor? Add the Adam MCP server to your client, sign in, and start Adam tasks from there. The CAD work still runs through Adam's harness and lands in your CAD documents; your client follows progress and answers questions when Adam needs a decision. The MCP is in beta.

Already on a ChatGPT plan? Connect your Codex subscription under Settings and Adam lists the models it exposes in a separate Codex section of the picker, billed against your ChatGPT allowance rather than your Adam usage.

Through the API, GPT-6 Astra lists at $10 per million input tokens and $50 per million output tokens, with a 1,050,000-token context window. Real CAD work is agentic: dozens of tool calls per part, each carrying the conversation and file context, so a single complex task can run into meaningful money on the raw API. CAD Arena's per-run cost in Fusion, about $5 for Astra against about $16 for Fable 5.1, gives a sense of scale for a single component.

Inside Adam, Astra draws from your plan's included usage, shown as a percentage on the Billing page, with a separate Ultra usage limit so one heavy session can't drain the whole allowance. Adam charges token usage at its underlying cost with no markup. See pricing for the current plans.

Comparing tools rather than models? See Adam vs Fusion MCP and Adam vs FeatureScript MCP.