I Gave GPT-6 Astra One Prompt and It Drew a 118 m² Floor Plan: Then It Asked Me for Money.

GPT-6 Astra had been out for a few days when I tried one thing with it: could I get a usable floor plan without opening a single piece of professional software — just one prompt?

I didn't go easy on it. The prompt ran close to 500 words and packed in everything about a real household: 118 m² gross, roughly 98 m² usable, three bedrooms, two living areas, two bathrooms, north at the top with the entry door on the north side, two adults plus a four-year-old and a cat, one adult working from home three days a week and needing somewhere quiet enough for video calls. Nine layout rules, from separating active and quiet zones and keeping the entry sightline off the bathroom, to 900 mm minimum clear width on main circulation, an enclosed kitchen because someone actually stir-fries in there, wet/dry separation in both bathrooms, and a service balcony holding a washer, a dryer, and the cat's litter box.

Here's the verdict up front: it was far better than I expected in ways I didn't expect, and it failed at the things I assumed it would obviously get right.

Step 1: The 2D Floor Plan — 2 Minutes 28 Seconds

That runtime is worth pausing on. This isn't instant image generation. It was thinking, and most of those two and a half minutes went into translating my nine constraints into actual spatial relationships.

What It Got Right

Orientation: all correct. Primary bedroom and living room face south, kitchen and child's room face north, entry door sits on the north side. The active/quiet split holds up — primary bedroom and child's room anchor the top and bottom of the west side, with the living and dining zone as one connected public block between them. You walk in and face the dining table, not a toilet or a headboard. The kitchen is properly enclosed, with a door. The balcony has the washer, the dryer, and the litter box, the last one even labeled "Cat Litter."

Dimension strings are clean and consistent. Total width reads 10500, and the 3400 / 4100 / 3000 segments add up.

This isn't a picture that merely resembles a floor plan. It genuinely laid out the space against the constraints I gave it.

save pin

Problem One: The Areas Don't Add Up

Every room carries a labeled area. I added them up:

Stated usable area: 98 m². That's an 8.3 m² gap. Corridors and internal walls do consume floor area — but there's barely any dedicated corridor in this plan; the public zone is one continuous block. The plan never accounts for the difference and never labels it.

The area figures look professional, but they were written on, not calculated.

Problem Two: Quiet Downgrades

The study was supposed to get a pull-out sofa bed. It drew a regular three-seater. Both bathrooms were supposed to have wet/dry separation; on the 2D plan there's no visible partition. I also specified a 2.9 m ceiling height, which a plan view can't express anyway — that one's on me.

Step 2: The 3D Render — 1 Minute 52 Seconds

I typed one line: "generate a 3D render." No new conditions.

Half a minute faster than the 2D pass, which tracks — the spatial relationships were already locked in, so this was just redrawing the same scheme from a different angle.

What Carried Over

It didn't hallucinate a new apartment. Wall runs, door swings, and furniture positions all line up with the 2D plan — this is the same scheme in a second representation.

Some things that were vague in 2D actually got resolved in 3D: the entry shoe cabinet and bench show up, and the glass shower partitions are visible. Wet/dry separation only really got delivered at this stage.

What Got Lost

Room labels and areas are gone — normal for a render, but it means neither image is complete on its own. The study still has a sofa, not a sofa bed. The primary bedroom closet was drawn as a walk-in in 2D and turns into a row of louvered doors in 3D; they don't match. The living room TV console disappeared and the TV became a floating black slab. The wall cut heights are inconsistent — the study wall on the east side sits noticeably lower than the primary bedroom wall, and the camera angle is a little off.

So the honest situation is this: you need both images together to reconstruct the full scheme. Neither one is usable on its own.

save pin

Step 3: Where the Free Path Ends

The 3D render got me a little carried away, so I tried to turn it into a rendered video.

It walked me through installing a video plugin. Installation done, and the message appeared: Generating video requires a paid plan.

Runway's annual pricing: Standard $12/mo for 625 credits, Pro $28/mo for 2,250 credits, Max $76/mo for 9,500 credits. The free plan doesn't include video generation at all.

You're Not Using One Tool

The first two steps happen inside the model. The third jumps straight to a third-party paid product. The experience feels continuous; the billing isn't. You think you're using ChatGPT, and you're being routed into signing up for Runway.

The Paywall Sits at Peak Sunk Cost

I'd already written a 500-word prompt, run two generations, and seen a result I was fairly happy with. That's exactly when it stopped me. Smart placement. Also deeply annoying.

save pin

So What Is It Actually Good For?

Use It For

1. Communicating intent. Explaining "the kind of home I want" to a designer or your family — one image beats ten minutes of talking.
2. Early-stage exploration. One prompt, one layout. Testing ten ideas costs almost nothing.
3. Content and visuals. Diagrams, deck slides, article images — more than good enough.

Don't Use It For

1. Anything requiring precision. The areas don't reconcile and the dimensions don't close. Take this to a quote, an order, or a job site and something will go wrong.
2. Anything requiring revision. What you get is an image, not a dataset. Want to shift the primary bedroom wall 300 mm to the east? You can only re-describe the whole thing and roll the dice again. You can't drag it. In a room planner, that same change is one drag and nothing else moves — which is the entire difference, and it has nothing to do with how good the AI is.

Every Pain Point in One Table

Six problems, two root causes: there's no underlying data, and the chain doesn't close.

Pictures Aren't Drawings

This is the part that actually matters: an AI-generated floor plan is a picture, not a drawing. A picture communicates intent. A drawing carries construction. What separates them isn't resolution — it's data structure.

Time cost magnifies the problem. One revision is five minutes. Ten revisions is close to an hour, and every single one is another roll of the dice — you can't change one thing, you can only re-describe the entire brief and hope it doesn't break what it got right last time. In a tool that stores the scheme as data, dragging that wall 300 mm takes three seconds and nothing else moves.

That's the real price difference between a picture and a drawing: not image quality, but the cost of every edit.

Astra's ability to take a long, messy, human brief and turn it into coherent spatial relationships is a genuine step forward — noticeably better than six months ago. But what it outputs still lives at the presentation layer. To go one step further — adjust dimensions, calculate materials, export to CAD, produce a quote — you still need a tool that stores the design as data.

If You Want to Try It Yourself

1. Be specific in the prompt. Don't write "three bedrooms." Write who lives there, how many people, whether there are kids, elderly, or pets, who works from home. It responds better to lifestyle constraints than to parameters.
2. Check the hard requirements yourself. It will quietly downgrade specifics — a sofa bed becomes a sofa, a partition just doesn't appear. Read your own brief back against the image after each generation.
3. Add the areas up manually. The numbers it prints aren't guaranteed to reconcile.
4. Run both 2D and 3D. They fill each other's gaps and cross-check each other — wherever the two images disagree is usually where it hadn't fully worked something out.
5. Budget the time. 2D took 2m28s, 3D took 1m52s. One complete pass is close to five minutes, and one revision means running the whole thing again.
6. Decide on paying only when you want video. The first two steps are free. The third is a different company's bill.

The Same Job, Two Pipelines

The 80-second difference isn't the point. The rows below the total are.

Astra runs three steps that produce one picture. A design tool runs three steps that produce one file — and the second and third steps can be re-run independently, because the first one left something behind to build on. That's why the second revision in Astra costs another four minutes and twenty seconds, and the second revision in a design tool costs seconds.

FAQ

Can GPT-6 Astra actually generate a floor plan?

It can, and the layout logic is sound. In this test it correctly handled orientation, active/quiet zoning, entry sightlines, and an enclosed kitchen — all from a written brief, with no software involved. What it produces is a rendered image of a floor plan, not a drawing file. It looks like a plan; it doesn't behave like one.

Are the dimensions and areas accurate?

The dimension strings were internally consistent — total width 10500 mm, with segments that add up. The areas were not. The ten labeled room areas summed to 89.7 m² against a stated 98 m² usable area, an 8.3 m² gap the plan never explains. Treat printed areas as decorative and add them up yourself.

Is it free?

The 2D plan and the 3D render are covered by a ChatGPT subscription. Turning the render into video is not — that step routes into Runway, whose annual plans start at $12/month and whose free tier excludes video generation entirely. The handoff is easy to miss because the experience feels continuous.

How long does it take to generate a floor plan?

2 minutes 28 seconds for the 2D plan, 1 minute 52 seconds for the 3D render from the same brief. One full pass is close to five minutes. That matters more than it sounds: since you can't edit the output, every revision costs another full pass.

Can I edit the floor plan it generates?

Not directly. The output is a flat image with no underlying geometry, so there's nothing to select or drag. Moving a single wall 300 mm means re-describing the entire brief and generating again — and the new version may break something the previous one got right. This is the core limitation, not a resolution problem.

Can I use an AI-generated floor plan for construction or permits?

No. The areas don't reconcile, the dimensions don't close as a set, and there's no layer structure, no wall types, no material data — none of what a construction document requires. Use it to communicate an idea, then have the actual drawing produced in a tool that stores the design as data.

Can it export to CAD or DWG?

No. Output is image only. If you need CAD, the realistic workflow is to use the generated plan as a reference and rebuild it in a design tool that exports drawing files.