industry

I think ChatGPT will eventually replace all CAD software

· 134 views

An opinion from the Nora3D team, illustrated by a Codex + Nora3D modeling session recorded on September 20, 2026.

I think ChatGPT will eventually replace all CAD software. The reason is simple: people want to describe a design, see a usable model, and change it without translating every decision into a sequence of commands.

That is a prediction about where design software is heading. Today's demonstration is smaller and more concrete: Codex creates an iPhone Duo exterior model through Nora3D, then changes its configuration after the user types “fold it.” Those two words are the most interesting part of the recording.

Watch the demo: from a prompt to “fold it”

A 53-second edit of a 26-minute screen recording. Waiting periods are removed and selected segments are accelerated, with speed labels. The video is silent and includes English scene titles.

The user starts by asking for a 3D model. In the connected Nora3D workspace, a solid appears, the second body half takes shape, and the model develops display surfaces and camera details. The user can inspect the result in the CAD viewport. A later message, “fold it,” requests a different configuration, and the geometry updates to a folded form.

Here is a text description of the edited video, including what is actually visible:

Video time Visible action What it demonstrates
00:00–00:07 Result preview, then the English modeling request The intended object is expressed in a prompt.
00:07–00:22 A solid appears; the body halves and display region develop The agent's tool calls produce changes in the CAD workspace.
00:22–00:36 Camera-side and unfolded views are inspected The result can be viewed from different directions.
00:36–00:45 The user sends “fold it”; the configuration changes A short follow-up acts on the existing model.
00:45–00:53 The folded exterior is shown The session ends with a visible revised configuration.

“iPhone Duo” is the name used in this demonstration. Treat the result as an exterior modeling exercise, not an official Apple CAD file or a verified reproduction of a shipping product. The recording does not establish dimensional accuracy, internal construction, hinge mechanics, or manufacturing readiness.

Why the follow-up matters more than the first model

An initial model is only a starting point. Real design work involves repeated changes: make the enclosure thinner, move a mounting point, add clearance, or try another configuration.

Each instruction depends on context. An agent needs to connect the user's words to the current document, identify the relevant geometry, perform an operation, and show what changed. “Fold it” is compelling because the user refers to an object that already exists. The conversation continues through the model.

This is where I expect the largest change in CAD interaction. More people will be able to explore a design without first learning the complete command vocabulary of a particular application. Experienced designers may spend less time expressing routine edits and more time evaluating the result.

The demo illustrates that direction. It is not a benchmark of success rates across arbitrary parts, and it does not establish that every short instruction will work.

What ChatGPT, Codex and Nora3D each do

These names describe different roles in the workflow. Keeping them distinct makes the prediction easier to evaluate.

  • ChatGPT represents the conversational experience at the center of this argument: explaining an intention and refining it through dialogue.
  • Codex is the agent used in the recording. It works with connected tools to carry out the modeling task.
  • Nora3D supplies the CAD workspace and modeling operations used in this demo. The visible geometry is created and changed there.
  • MCP, the Model Context Protocol, is the connection mechanism through which compatible agents can access external tools. OpenAI documents this capability in its Codex MCP documentation.

The workflow can be described as a design request → agent tool calls → CAD operations → a visible result → a follow-up request. Geometry software still has a job to do. An agent gives the user another way to direct it.

For the setup path, see Nora3D's Codex integration guide. For the modeling concept, start with text-to-CAD.

What I mean by “replace all CAD software”

The strongest version of the prediction is about which interface people choose to open when they want to design something. I expect a conversational agent to become that entry point for an increasing share of work.

A designer could describe an enclosure, inspect the first version, ask for several changes, and hand the result to a specialist workflow. The underlying systems might still contain solid-modeling kernels, constraint solvers, simulation tools and manufacturing software. The user would direct more of that work through a shared conversation.

That is the future Nora3D is working toward. A request should remain connected to the actual design, and a revision should produce geometry the user can inspect.

“All” is deliberately ambitious. This session does not prove that every professional discipline will move at the same speed, that any vendor will disappear, or that the transition has a known deadline. Complex assemblies, drawings, simulation, manufacturing and controlled releases will demand much more than a convincing exterior model.

The tests an AI CAD workflow still needs to pass

The next useful question is how to tell whether a generated model can be trusted for its intended use. I would evaluate an agent-driven CAD system against five concrete checks:

  1. Requirements: Do measured dimensions, clearances and feature locations match the brief? A plausible silhouette is insufficient.
  2. Targeted changes: Does a request affect the intended body or feature while preserving the rest of the design?
  3. Geometry quality: Is the result valid for the next operation, and can problems be identified before handoff?
  4. Persistence and exchange: Can the model be saved, reopened and checked in the downstream tool? An export needs its own validation.
  5. Engineering review: Have materials, tolerances, loads, assembly and manufacturing constraints been evaluated for the actual application?

Nora3D's guide to editable CAD, B-Rep and feature history explains why editable solid geometry and a native parametric history are different things. A successful STEP handoff also needs more evidence than a screenshot. For a printed part, follow the separate CAD-to-3D-printing review process.

Those checks are a practical development target. They help distinguish a promising demonstration from a dependable engineering workflow.

Try a small, reviewable modeling task

Start with a simple object and a measurable brief. For example: “Create a rectangular enclosure, 80 × 50 × 25 mm, with 2 mm walls and an open top.” That is an illustrative prompt, not a tested result from this recording.

Inspect the geometry, then request one change: “Increase the wall thickness to 3 mm while keeping the outside dimensions unchanged.” Check both requirements after the edit. This makes it easier to judge whether the workflow understands your intent.

You can open Nora3D or follow the Codex connection guide. Begin with a disposable test document and review each meaningful change before applying the same approach to an important design.

Frequently asked questions

Can ChatGPT create CAD models today?

An AI agent can create or modify CAD models when it has a compatible modeling tool connection. This recording specifically shows Codex using Nora3D. A text response or a rendered image alone does not establish that an editable CAD model exists.

Is this ordinary ChatGPT working without CAD software?

No. The demonstration uses Codex with Nora3D. Nora3D provides the CAD workspace and modeling tools. The broader claim about ChatGPT replacing CAD interfaces is a prediction, not a description of standalone capability proven here.

Does the demo show a working folding mechanism?

It shows the exterior geometry changing from an unfolded to a folded configuration. It does not establish a mechanically validated hinge, a motion simulation, or collision-free movement through every intermediate angle.

Was the complete model made in 53 seconds?

No. The source recording lasts about 26 minutes. The published clip removes waiting periods, selects the key changes and accelerates some segments. Its duration is not the modeling time.

Can this model be manufactured directly?

The video does not provide that evidence. It shows an exterior modeling exercise. Manufacturing would require verified dimensions, materials, tolerances, internal design and process-specific checks.

What does this mean for professional CAD users?

It suggests that more modeling work may begin with a request and continue through dialogue. Professional judgment remains essential for defining requirements, examining geometry and validating a design. This session offers an example of the interaction, not a replacement for that validation.

The iPhone Duo session is a beginning: describe an object, inspect it, then ask for a change. If that loop becomes reliable across more of engineering, the way people use CAD will change profoundly. That is why I think ChatGPT will eventually replace all CAD software as the primary interface for design.