Direct answer: GPT-6 Astra is unlikely to replace your CAD software. It is far more likely to sit above it, reading instructions, inspecting drawings, writing code and running tasks while the CAD system keeps hold of the geometry. That is a genuine shift. It is also a long way from something we would let near a production drawing without a person checking the output.
We have been supplying and supporting CAD software for over twenty years. Our technical team have spent that time installing it, fixing it, and cleaning up drawings that were technically correct and practically unusable. That experience is why we read a benchmark score carefully rather than excitedly.
Whatever AI does next, it runs on top of a CAD platform.
BricsCAD is native DWG and has the open API this article describes. Try the full Ultimate package free for 30 days.
CAD has become enormously more powerful over the last three decades without ever really changing the thing that sits between you and the software. You pick the command, select the geometry, type the value, and repeat until the drawing matches what was in your head when you started. Faster hardware and better tools have made each of those steps quicker, but the loop itself has not moved.
Astra could start to move it.
OpenAI released the model on 3 September 2026 and describes it as state of the art on computer use, browsing, software engineering and professional work, capable of operating desktop software and of using functions, file search, web search and computer use together through the API (OpenAI). For CAD that combination matters more than any chat interface, because a model that can read a task, inspect a drawing, write code, run the software and then check its own result is doing something categorically different from answering questions.
The number that got CAD people’s attention was BenchCAD. Astra scored 95.9 per cent geometric overlap, against 83.3 per cent for GPT-5.6 Sol and 84.3 per cent for Claude Fable 5.1.
That figure needs a qualifier, and the qualifier is the interesting part.
The 95.9 per cent was achieved in the agentic setting, where the model is given a Python sandbox and can render the part, measure it, and try again before submitting an answer (BenchCAD). In the plain setting with no tools, Astra has published no figure at all, and GPT-5.6 Sol still leads at 0.706. OpenAI also notes that the Claude comparison runs used modified evaluation settings.
So the honest version is this. A model that can render, measure and iterate scored 95.9 per cent on shape reconstruction. That is a real result and a real step forward. It is not the same claim as “AI can now do CAD”, and the difference between those two sentences is the whole subject of this article.
Worth noting alongside it: on the Artificial Analysis Intelligence Index v4.1.1, an aggregate run by a third party rather than by a vendor, Astra scores 61.2, behind Claude Fable 5.1 at 65.7 and Claude Opus 5 at 63.1. Launch tables are written by the people launching.
The interesting shift is not AI generating finished designs. It is the move from telling the software how to do something to telling it what you want.
Today a drafter picks a layer, starts the dimension tool, selects the geometry, places the dimension, adjusts its settings, and then does the same thing another forty times across the sheet. With a working natural language layer, the same person might instead say:
Check this drawing for missing dimensions and add the ones needed for manufacture. Do not change any existing dimensions, and show me what you plan to add before you add it.
The first version describes a procedure. The second describes an outcome, and leaves the AI to work backwards to the steps. That is the whole shift in one example, and it is why the “show me first” clause matters as much as the instruction itself.
This is not speculative. BricsCAD, Autodesk, Siemens and SOLIDWORKS are all adding natural language features and AI assistants to their products already. What a model like Astra could add on top is the ability to hold a longer chain of work together.
Either way, the CAD application keeps hold of the things that matter: exact geometry, constraints, object data, APIs and file formats. The AI handles planning.
If you want to know where AI lands first, look at 2D drafting, because a large share of that work is rule-following, error-hunting and repetitive editing. Those are exactly the tasks a language model is good at.
Ask any CAD technician what actually eats their week and you will not hear “designing things”. You will hear about chasing text that is the wrong height, hunting down the one overridden dimension somebody fudged three revisions ago, checking a drawing against a standard nobody has written down properly, and updating the same note across forty sheets because a client changed their address.
That is the work. It is also precisely the work an assistant could take on:
Scripts already do some of this, but scripts are rigid. They run one fixed sequence and fall over the moment the drawing is not quite what they expected. An AI layer can cope with variation, which is what makes it useful for annotation, blocks, title blocks, standards checking and drawing cleanup.
Unglamorous, and worth far more than any demo of a robot designing a bracket.
3D is harder, and the gap between what looks impressive and what is actually useful is much wider.
A rendered image is not a CAD model. A real model carries sketches, dimensions, constraints, features and the relationships between them, and it is those relationships that let somebody change it in six months without starting again. This is why BenchCAD is a more interesting test than it first appears: it asks for executable parametric CAD, not a picture.
Consider a request like this one:
Increase the width of this mounting plate by 20 mm, keep all four holes symmetric about the centreline, and do not take edge distance below 12 mm.
To do that properly the system has to read the feature tree, find the right parameters, understand which rules are not allowed to break, make the change, and then verify it held. Get any one of those wrong and you have a plate that looks right and fails inspection.
Astra’s coding ability points at a different route entirely. Most CAD systems expose scripts and APIs, so a model does not need to push a mouse around at all. It can write code that calls the CAD system’s own functions, which is both more reliable and much easier to audit afterwards.
That turns natural language into a front end for CAD automation. You give the intent, the AI plans, the CAD system builds the geometry, the AI checks the result, and you approve it.
No, and we would say that even if we were not in the business of selling CAD software.
The reason is not sentimental. It is that the current generation of models is bad at exactly the part of the job that makes someone a designer rather than an operator.
The clearest evidence comes from neuralCAD-Edit, a 2026 benchmark built by recording ten professional designers editing real models in real CAD software while talking, pointing and drawing. When frontier models were measured against those experts, the best of them scored 53 percentage points below the humans in acceptance testing (Perrett et al., arXiv:2604.16170). Astra was released after that study, so it is not an Astra result, but the size of the gap tells you where the problem sits.
A note on names, because they are confusingly similar: neuralCAD-Edit is an academic benchmark. Autodesk’s Neural CAD is a commercial product line. They are unrelated.
The BenchCAD researchers found the same pattern from the other direction. Frontier models regularly got the broad shape right and then failed on the parametric structure underneath, with missing detail, misread engineering parameters and wrong modelling sequences (BenchCAD, arXiv:2605.10865).
One example from that paper is worth more than any benchmark table. A model produced a part scoring 96.1 per cent geometric similarity. The requested modification had not actually been made, because a cylindrical boss added afterwards partly closed the opening again. Ninety-six per cent similar. Wrong part.
That is the difference between similar in shape and correct as engineering, and it does not show up in a screenshot. It shows up on the shop floor.
What does change is the balance of the job. If an assistant handles standards checking, annotation and revision comparison, then less of the week goes on documentation and more of it goes on decisions, trade-offs and checking. For anyone who already knows what they are doing, that is a better job, not a smaller one.
The people who should be paying attention are the ones whose value is mostly speed at repetitive drafting. That work will get automated, and it has been getting automated steadily for twenty years. The answer is the same as it has always been: get better at the judgement part.
The sensible architecture is a layer, not a replacement. The CAD system stays in charge of geometry, constraints, drawings, assemblies and file formats. The model handles planning and orchestration.
OpenAI’s computer use tooling lets a model read screenshots and request actions through code, keyboard and mouse. That is impressive, and for serious CAD work it is the wrong approach. Direct API connections beat screen clicking every time, because they are faster, they fail predictably, and you can see afterwards exactly what was called.
Done that way, an assistant could answer questions like:
Find every 8 mm hole in this model and tell me which ones are not part of a standard feature.
A modern API can expose a set of safe, named actions instead of handing over control of the whole application. Read drawing properties, query layers, create dimensions, change parameters, generate views, run standards checks. Each one is a door you can open or leave shut.
This is the route to something dependable. Language and planning for the ambiguous parts, the CAD kernel for the exact ones. It also means somebody still has to decide what the assistant is permitted to do, which standards apply, and whether the result is good engineering. That somebody is an engineer.
Some of this exists today, which is worth knowing before anyone reorganises a workflow around a model released last week.
BricsCAD V26.2 added BricsCAD Assist, an AI panel inside the interface that answers questions about commands and workflows without sending you off to a browser. AI-Predict learns how you personally use the software and suggests the commands you are likely to need next, in the ribbon and in command line auto-complete.
Health Advisor is the one that fits everything above most closely. It runs drawing health checks in the background on every drawing you open and reports what could be improved. It changes nothing on its own. It tells you, and you decide.
That is the inspectable, reversible, human-approved pattern this whole article has been arguing for, and it is already in the product rather than in a research paper.
Health Advisor, BricsCAD Assist and AI-Predict are in the current release.
Try them on your own drawings. Full BricsCAD Ultimate, free for 30 days, no payment details.
This is the question that stops most firms, and rightly. CAD files carry intellectual property, unreleased product data and client information, and quite a few of our customers work under agreements that say precisely where that data is allowed to sit.
OpenAI says business data sent through its business products and API is not used to train models by default, and that some API customers can use Zero Data Retention.
That answers one question and leaves the harder ones open. Not just can the AI read our files, but should it, where does the data physically go, what is it permitted to change, and who signs off the result? Those are policy decisions, and most firms have not made them yet.
If you are the person who has been asked to work this out, it is worth writing it down before anyone installs anything.
Broadly yes, but not in the form the headlines suggest, and not this year.
The strongest signal is not Astra on its own. It is that several things are moving at once: natural language interfaces appearing in mainstream CAD products, AI-assisted modelling maturing, and research that tests whether general models can build genuinely editable CAD structures rather than convincing pictures.
Commands are not going anywhere. Menus, scripts, direct modelling, parametric features and APIs will all still be there in ten years. Natural language becomes another way in, not the only one.
Instead of:
LINE. Pick first point. Pick second point.
You might say:
Bring this drawing in line with our CAD standard and show me every change before you apply it.
Traditional CAD makes you translate intent into commands. AI-assisted CAD does more of that translation for you. For an experienced user that is not less control. It is less time spent on the routine steps and more spent on the parts that need judgement.
Which is why we find Astra interesting without being especially excited. Not a replacement for CAD, not a replacement for the designer, but a new layer between what you want and how the software gets there.
For 2D drafting and 3D modelling alike, that is a bigger change than another button on the ribbon.
Not on current evidence. In the neuralCAD-Edit benchmark, where frontier models were measured against professional CAD engineers on realistic edits, the best model scored 53 percentage points below the human experts in acceptance testing. What changes is the mix of the job: less time on repetitive documentation, more on design decisions and checking.
Getting better at the judgement part is a training problem.
Our Essentials Course is taught by people who have spent twenty years in real drawings, not in slide decks. It covers the fundamentals that decide whether you can tell a good AI output from a plausible one.
No. It is far more likely to sit above CAD as a reasoning and orchestration layer. The CAD application still handles geometry, constraints, dimensions, object data, the geometric kernel, APIs and file formats, while the model interprets intent and plans the work.
95.9 per cent geometric overlap on BenchCAD, against 83.3 per cent for GPT-5.6 Sol. That result was achieved in the agentic setting, where the model can render, measure and iterate using a Python sandbox before answering. No score has been published for Astra without tools.
It can produce something that looks like a CAD drawing. Producing one that is properly constrained, correctly dimensioned and editable afterwards is a much harder problem. One BenchCAD example showed a part scoring 96.1 per cent geometric similarity where the requested change had not actually been made correctly.
You describe the result you want rather than every step to get there. Instead of selecting objects, changing layers and applying dimensions one at a time, you ask for a drawing to be brought in line with a standard, review the proposed changes, and approve the ones you want.
Because professional models are not just shapes. They depend on sketches, dimensions, constraints, features and relationships that allow reliable editing later. A model that only produces a convincing 3D object is much less useful than one that produces structured parametric CAD with the design intent intact.
Wrong geometry, misread design intent, poor modelling choices, missed manufacturing requirements, data and IP exposure, and too much trust placed in unsupervised changes. Changes should stay inspectable, reversible and subject to human approval.
Everyday 2D drafting, documentation and quality checking. Standards checks, layer mapping, annotation review, finding overridden dimensions, comparing revisions, updating title blocks and flagging exceptions. Repetitive multi-step work where the rules are known but the drawings vary.
It depends entirely on which tool and which account type. Consumer and business tiers have different data handling, and CAD files often carry client IP or contractual restrictions on where data may be processed. Check the terms and write an internal policy before anyone starts uploading drawings.
Yes. The parts of the job being automated are the repetitive ones. Understanding geometry, tolerances, standards and manufacturing constraints is what lets you tell a good AI output from a plausible one, and that is becoming more valuable rather than less.
