Can AI do 3D modeling?

A real brief, not a test prompt

We tried to simulate a real-world scenario: the same brief a client would send us, or rather a similar brief a client already sent us. Earlier this year our team built the interactive configurator for the GRAU Salt, a cordless aluminium lamp in eight colourways from GRAU, a German lighting brand that sits closer to art than to hardware. Roughly 20 hours combined between three people: 8 of modelling, 4 of design, 8 of development, across a few days, with feedback from the client and a few iterations in between. It is live and the client approved it.

So we handed a similar job to Fable 5.1 and GPT-6 Astra. Each got the manufacturer's own CAD, the DWG and 3DS that GRAU publishes, and a link to the live product page. Then the whole job: a web-ready asset and a configurator built to sell. Eight colourways, a night mode, measurable dimensions, hotspots written from the brand's own copy, working on a phone.

They chose their own tools. Nobody helped once the clock started. Our own configurator was named and put off limits.

Both of them shipped

Not the demo. Two working product pages, on the brand's colours and type, the lamp in the middle, the features annotated. The goal was the full job, from zero to done, and both got there. One of them added three light temperatures nobody asked for, at the correct kelvin values from the brand's own page.

The night mode is a nice effect in both, and it is worth saying plainly: it is far from realistic. Light that behaves like the real lamp is very hard to generate, and neither build gets there.

Neither of them touched the geometry

The usual thinking is that CAD is heavy and must be reduced for the web. In this case both recognised the file was already right for the job and that there was no need to change the geometry. They left every vertex where the manufacturer put it. The right call, and not the obvious one.

We did not take their word for it. Every vertex of both assets was re-measured against the original CAD, and the results were checked again by a second pass. Astra: 0.01 µm of drift (a hundredth of a micrometre; a micrometre is a thousandth of a millimetre). Fable: 9.7 µm. Both reports held up.

Here is where it gets interesting

Same geometry, different outputs. Fable optimised for weight and loading speed: 16-bit positions and meshopt compression, a file 2.8 times lighter, with under ten microns of drift. Astra did not compromise on exactness: no quantisation, no lossy compression, zero drift, and a file passed through the Khronos validator with no warnings.

In this case both are fine, because the product itself is tiny. Fable's logic, get the download as small as it can be, is good. Astra's logic, never give up quality, is good. Two excellent answers to the same question.

Materials: this is the hard part

Colour and finish are the hardest part of this work. It is where our own team spent quite a few iterations to get the materials right, because the client is in love with their colours and holds them to a high standard. Powder coat, brushed aluminium and an opal diffuser do not come from a hex code.

Neither model got close. The shapes are identical; the surfaces are not. Look down each column: the same colour name, three different objects.

Sand White

Polished

Blood Orange

Raw

Chocolate

Black

Sun Orange

Ultramarine

Ours, approved by the client

GPT-6 Astra

Fable 5.1

Small details, big difference

The charging port is where the two builds part ways. Astra modelled it as an opening in the base, with the indicator LED next to it, the way it is on the lamp. Fable's port reads as a flat pale mark that you could mistake for a sticker. Missing a detail versus not missing it is a big deal on a product page, because the customer zooms in on exactly these things.

And on the Astra build, the Ultra Touch Dimmer hotspot flickers at certain angles. A small CSS mistake, and exactly what a client notices first.

What it took

We are not going to disclose what our team costs. But this is roughly the time it took, side by side. Hours as they were actually spent, across a few days, with a client review in the middle.

The scorecard

These are our scores. Two rows are measured; the rest is our judgement, and we would rather say so than dress it up as a benchmark. We are not scoring our own work. It is paid client work, delivered and approved, and that column says exactly that.

Trust is the row that matters if you work in 3D. Not that a model made a mistake, but that you have to check every detail to find it, and once you find it you either fix it by hand or reprompt. Both cost time you did not plan for.

What this does not prove

1. ·This is an isolated situation. One product, one run each. We will do one type of product every week to see where things stand.
2. ·This lamp does not have many details to cover. With more complex models we expect many more mistakes.
3. ·The CAD existed. A brand with four phone photos and a shop link is a different customer, and half of what a studio does is absorbing that.
4. ·We judged the AI builds ourselves. Read it knowing who wrote it.