A Fable class model, as Simon Willison uses the phrase in 2026 in LLMs (so far), is a model that will solve your problem effectively through brute force provided three things are in place: the goal for what you want to build is clearly defined, the constraints around that goal are given as unambiguous instructions, and the model has access to the necessary tools to achieve it. He names the class after Claude Fable 5, released in June 2026, which he calls the first public glimpse of such a model, and counts GPT-6 Astra and GPT-5.6 among others he places in it.
Usage
Willison draws two readings from the idea. On one hand it looks like a direct threat to software engineers, because it means models can build effectively any piece of software that can be defined in this way. On the other, he argues, defining goals, providing unambiguous instructions and working out the right tools are “kind of what software engineering is”: doing them well takes a lot of experience and skill, and someone who can do them well now has, in his word, superpowers. He says this realisation — that there is still a lot of skill in driving models this capable — helped him somewhat with the feeling he calls Deep Blue.
The category is his own informal one. His placement of particular models in it rests on his own use and comparisons, such as calling GPT-5.6 “definitely a Fable class model” while judging it possibly not quite as good as Fable.