Key Takeaways

  • Claire Vo found that Anthropic's Fable 5.2 fails at autonomous front-end web design, producing visual layouts with what she called "incredibly concerning taste."
  • The model follows strict design systems and style guides accurately, but its aesthetic output degrades the moment prompts leave layout decisions unconstrained.
  • Anthropic does not provide a native image generation engine, forcing builders to route visual asset creation out to OpenAI Image Gen 2.5.
  • For software builds, Vo separates model roles by pairing Opus 5.5 on front-end design with Fable 5.2 handling back-end architecture.
  • Despite handling complex tasks like 3D ambient typewriters and flat-pack assembly from photos, Fable 5.2 cannot be trusted with raw visual aesthetics.

Fable 5.2 Needs a Style Guide to Build Usable Web Pages

Claire Vo entered her evaluation of Anthropic's Fable 5.2 as an established Claude skeptic. The model won her over on raw speed, back-end architecture, and creative visual builds like an animated marketing video and a custom 3D ambient typewriter. It could even parse photos to guide flat-pack furniture assembly. Yet the moment she asked it to write front-end code without explicit constraints, the output collapsed.

“It's bad at web design. It's like real bad at web design,” (play from 16:54) Vo said. Her testing showed a clear dividing line between rule execution and aesthetic judgment. When given a detailed brand manual or an exact design system, Fable 5.2 implemented the specs cleanly. But when handed an open prompt to design an interface from scratch, its visual decisions were clumsy. “It's really tasteful when you give it a good brand or style guide and it's good at following it, but left to its own devices, it has incredibly concerning taste. Like she just does not have good taste.” (play from 6:13)

The model writes functioning HTML and CSS without syntax errors. The real breakdown is visual judgment. Fable 5.2 lacks the default visual intuition that developers expect from a top frontier model. It will position elements awkwardly, select unpolished typography pairings, and assemble layouts that look like early draft software from a decade ago. If you expect the model to replace a designer or invent a visual language on the fly, you will spend your time fixing its mistakes.

Anthropic Still Lacks a Native Image Generation Engine

The aesthetic shortfall extends directly to Anthropic's product boundary. Unlike competitors that bundle diffusion models into their interfaces, Anthropic provides no native visual creation tool. That omission leaves a clear hole during creative development.

“And then I think this comes from a lack of image gen model really shows because when a model is really creative, you want it to be creative along a lot of aspects,” (play from 6:26) Vo explained. To get complete prototypes off the ground, she had to bounce between providers: “And I found myself a lot of times invoking OpenAI's image gen 25 to create images. And then you know Fable wasn't very good at that in particular.” (play from 6:37)

This limitation shapes how engineers must structure their AI toolchains. Instead of relying on a single frontier model to generate both code logic and assets, builders must split their stack. Vo's practical response is division of labor across model boundaries: “So, I might pair Opus 55 with Fable on the back end for some of the design stuff, but I really don't love the front end design of Fable.” (play from 17:33)

Opus 5.5 handles front-end layout decisions and styling where Fable falls short. OpenAI Image Gen 2.5 handles asset creation. Fable 5.2 stays focused on rapid back-end logic, architectural scaffolding, and functional code generation.

What to Do With This

Never ask Fable 5.2 to design an interface from an open prompt. Feed it an exported CSS design system with explicit color tokens, spacing scales, and typography components before asking for UI code, or hand the front-end layout tasks entirely to Opus 5.5. When your prototype requires visual media, route asset calls directly to OpenAI Image Gen 2.5 instead of expecting Anthropic's model to bridge the gap.