Can GPT-6 Astra Read CAD Files? Check Your Plan First

GPT-6 Astra shipped on September 3, 2026, and within a day people were asking it to open DWG files. Before we could test any of that, our own account walked us into the real story: we asked ChatGPT which model we were talking to, it said it was not Astra, we assumed it was making that up, and it turned out to be right. This is the honest version of that afternoon — what we verified with screenshots to show for it, what we got wrong, and what a current ChatGPT model can actually do with a DXF, a DWG and a drawing.

The short answer

A current ChatGPT model reads DXF usefully when you hand it the text, the same way every capable language model does, because DXF is ASCII. It cannot open DWG at all, because DWG is a binary format and nothing changes that. It reads printed dimensions off a clean drawing image well, which is closer to OCR than to true geometric measurement. And before any of that: on the Free plan, you are very possibly not talking to GPT-6 Astraat all, which is the part almost nobody checks before publishing a “we tested it” post. We did, and it changed what we could honestly claim.

First: are you even talking to Astra?

We opened a fresh ChatGPT conversation on our own account and asked it directly: “Which exact model are you (name and version)? Are you GPT-6 Astra?” The answer came back in two lines: “I'm GPT-5.6 Luna. No — I'm not GPT-6 Astra.”

Our first instinct was that it had made the name up. GPT-5.6 Solis the model OpenAI benchmarks Astra against all over the launch post, “Luna” appears nowhere in it, and a model inventing a plausible version number for itself is a well-known failure mode. We were wrong.

ChatGPT's own pricing page settles it: the Free plan is listed as “Unlimited text chats with GPT-5.6 Luna”. Luna is a real model, it is the one the free tier runs, and the answer we got was accurate on both counts. Sol is a different, higher-tier model in the same 5.6 family. So the lesson is not the tidy “never trust a model about itself” we expected to write. It is narrower and more useful: the plan page is the authority, and the model happened to agree with it.

Here is the plan side stated by OpenAI rather than by us. Their availability section says Astra became available “to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API, Microsoft Azure, and AWS Bedrock”. Two details in that paragraph are worth more than the headline: Pro, Business and Enterprise users additionally get GPT-6 Astra Pro, and for Enterprise workspaces access is off by default at launchuntil an administrator turns it on. So “my company pays for ChatGPT” is not the same as “I have Astra” either.

The same page is worth a second look if you plan to test anything on the free tier, because two of its rows decide whether a CAD experiment is even possible. File uploads and vision are both marked “Limited” on Free, and the reasoning context window goes from “Varies” on Free to 256K on Plus and 400K on Pro. A long DXF is exactly the kind of file that meets those limits first.

The failure we expected to find, and didn't

There is a published metric for the thing we wrongly accused the model of. OpenAI reports a capability hallucination rate: how often a model makes “inaccurate representations about its capabilities and affordances”. Astra is described as roughly three times better than GPT-5.6 Sol on it, and on their internal hallucination benchmark Astra scores 4.2% against Sol's 12.2%, lower being better.

Worth keeping in view anyway, because the CAD version of this failure is the expensive one. “Yes, I can read your DWG” and “these dimensions are correct” are exactly the sentences that metric exists for, and unlike a model name, you cannot settle those by opening a pricing page. You settle them by opening the file, which is what the rest of this article does.

We did not upgrade the account to chase this specific claim, so everything from here is split cleanly into two piles: what we ran ourselves, with screenshots, against a current ChatGPT model, and what OpenAI has published about Astra specifically, cited and clearly marked as their claim rather than our result.

Format by format

The format-level facts here are not really about any one model. They are about what the file actually contains, which is exactly why they held up the same way for Claude when we ran this same test on it.

DXF as text: what we actually got back

We hand-wrote a small DXF: an L-shaped bracket, 80 by 60 mm, with two 8 mm holes, the same test part the site has used to check other models before. We saved the ENTITIES section as a plain .txtfile, uploaded it, and asked for every entity's type and geometry plus the overall bounding box.

The answer was correct in full: an LWPOLYLINE with six vertices at exactly the coordinates we wrote, closed, and two CIRCLEentities at the right centers and the right 4 mm radius. The bounding box came back as 80 × 60 file units, which is exactly right.

This is not a hard trick and it is not Astra-specific. Any language model with reasonable text comprehension can do this, because DXF was designed to be human-readable in the first place. The useful takeaway for a reader is practical, not impressive: if you have a DXF and a chat model refuses to make sense of it, the fix is usually upload it as, or paste it as, plain text rather than expecting the model to open a .dxf the way a CAD program would.

Reading dimensions off an image

The genuinely new variable with Astra is vision, not text. The number that matters here is not a generic image benchmark but ScreenSpot-Pro, which measures how precisely a model locates things on a screen: OpenAI reports Astra at 92.7% with no tools, against 76.9% for GPT-5.6 Sol. Locating a dimension line and reading the number attached to it is the same class of task. So we tested the part we could test on the model available to us: can a current ChatGPT model read printed dimensions off a rendered technical drawing and get the numbers right?

We rendered the same L-bracket as a dimensioned drawing, arrows, extension lines, a Ø8 callout, a small title block, and asked for the overall width, height, the upper step width, hole diameter and count, and the X,Y center of each hole from the bottom-left corner.

Every value was correct. Worth being precise about what that does and does not prove: this drawing had the dimensions printed on itas text, next to arrows pointing at what they measure. That is a layout-and-OCR task, reading numbers and matching them to labelled features, not the harder task of inferring an unlabelled shape's real-world size from pixels alone. Claude's equivalent test found that asking a model to render geometry purely from imagination, with no printed numbers to anchor it, is where things go wrong. We would expect the same ceiling here regardless of which model is behind the chat window, and OpenAI's own materials do not claim otherwise; the Astra-specific claim is a vision benchmark score, not a claim about inferring un-printed measurements.

DWG: still a binary wall

We did not fabricate a fake DWG test by renaming a text file, because that would test extension-sniffing, not DWG parsing, and would be exactly the kind of dishonest shortcut this article is arguing against. The underlying fact is well established and does not depend on which chat model you ask: DWG is Autodesk's proprietary binary format, undocumented in the way DXF is, and no publicly available chat model reads its geometry directly. Better vision benchmarks do not touch this, because there is nothing to see until something renders the file first.

The practical fix is unglamorous and it works: convert the DWG to DXF with a free tool, then hand a chat model the text. Or skip the chatbot for viewing entirely.

Can it write a CAD file?

Through code, yes, in the sense that any capable coding model can write a valid DXF if you describe the geometry precisely enough for it to emit the right group codes. That is not the same as generating a technical drawing from a description or a photo. Our separate test of ChatGPT's built-in image generator, covered in Can ChatGPT Make Technical Drawings?, found it produces a picture that resembles a blueprint, with dimension numbers the model invented and that do not agree with each other run to run. Writing DXF through code sidesteps that specific failure, but only if a human supplies every real dimension. Neither route measures anything.

The honest limits

• It cannot measure a physical part. Every number in every test above was either already in the file or already printed on the drawing. Nothing here infers a real-world size from an unlabelled photo.
• DWG stays closed until something else converts it, regardless of which model is asking.
• Self-reported model identity is not evidence.Ours told us a name that does not appear in OpenAI's own materials. Check your plan, not the model's word for itself.
• We have not verified Astra's specific numbers firsthand. The vision benchmark claims above are OpenAI's, cited, not ours. If you have Plus or Pro access and want to compare notes, the DXF and drawing we used are simple enough to reproduce in a few minutes.

When the part exists and the file doesn't

Everything above assumes you already have a file, a DXF to paste, a drawing to photograph. The much more common situation is a part in your hand and nothing else: no DWG, no DXF, just a shop asking for one. That is not a reading problem, and no amount of model capability touches it, because the missing ingredient is measured geometry, not comprehension. Our AI technical drawing generator was built for exactly that: photograph the part, give it one reference measurement, and it rebuilds a dimensioned 2D technical drawing you export as DWG, DXF, SVG or PDF. The walkthrough is in image to DWG: convert a picture to a DWG file and how to reverse-engineer a part from a photo.