15 Iterations With Fable 5.1. One With GPT-6 Astra.

You can identify AI-generated web design in about a second. Gradient hero, a section that could belong to any company, buttons that look the same whether the brand is a bank or a skate shop.

That changed at launch, and OpenAI shipped it as a stated capability rather than a side effect.

Fifteen iterations, compressed into the first draft

The most concrete independent result came from MindStudio, who had built a site with Claude Fable 5.1 over roughly 15 iterations to reach a polished, on-brand state. Astra's first draft of a comparable site matched that refined version's feel with no refinement cycles. They tested seven one-shot sites in total.

Their explanation is spatial rather than stylistic. Astra puts elements on distinct planes: "a bike sits on one layer, rocks behind it on another, background imagery further back, and text on top of all of it." Scroll, and the planes move at different rates.

That is why the slop signature was always flatness. A model with no spatial model of the page can only stack blocks vertically, everything lands on one plane, and no amount of gradient fixes it. Depth is what those fifteen iterations were buying.

A 3D iPod that browses your agent threads

The single best interface demo of the week is also the most absurd, and the prompt is worth reading in full before you watch it.

Three things at once: a modelled object, a native application, and a faithful reproduction of a specific twenty-year-old interaction model. Generic models are worst at exactly that last part, because it requires knowing what a click wheel felt like.

Claire Vo got the same shape of result independently: a native Mac desktop app wrapping her Codex threads in a 1990s AOL Instant Messenger client, in a single shot.

The least glamorous demo, and the one with money attached

Nothing in this series is less exciting to watch and more relevant to a company that ships software. The roundups clocked the loop at 36 seconds.

It also operates the design tools

Claire Vo reported the same in Figma, where Astra assembled a finished YouTube thumbnail from generated assets rather than handing her the pieces.

The honest limits

The comparison that produced the 15-to-1 result also declined to call Astra better overall. Neither model got controlled, identical-prompt testing, and one Astra output shipped with an incorrect domain name rendered on screen.

So the claim is narrow and still large: the first draft got dramatically better. Nobody has shown that the twentieth draft did.

The operator read

If your team ships interfaces, the bottleneck moved.

It used to sit at production. Getting from an idea to something you could look at was slow, so you argued in Figma to avoid wasting build time. When the first draft is close, the expensive step is no longer making it. It is deciding whether it is right.

Which is Hiten Shah's point from the same week: "the prompt is not the skill. Knowing what good looks like is the skill." A team with taste just got much faster. A team without it now produces plausible work at speed, which is worse than being slow.

Part 2 of Astra, Week One.

John Tan

Founder and CEO of nativefirst.ai. Embeds with scaling founders and CEOs to ship Level-3 agents and AI workflows in production.