Tested: Claude Fable 5 sets a new bar for AI models, but takes control away from the user

Anthropic’s new Claude Fable 5, the first model in the so-called Mythos class, significantly outperforms all previously available AI models across a wide range of tasks. Ethan Mollick writes for One Useful Thing that the model can work autonomously for up to twelve hours, coordinate multiple AI agents, and deliver finished, complex results from just a brief instruction.

Mollick, who received early access to the model, tested it on tasks ranging from game development to academic research tools. In one example, he asked Fable to build an isochrone map showing real travel times between cities by plane, train, and road. The model independently launched multiple subordinate AI agents to gather over 2,200 flight connections, international rail schedules, and country-level road speed data. It then coded, tested, and refined the result, largely without human input.

A studio, not an assistant

In a more ambitious project, Mollick asked Fable to design and build a piece of research software called Concord, which calibrates human and AI judgements on qualitative data. The model worked for nine and a half hours and produced a fully functional tool that, according to Mollick, researchers have needed for years but that was never commercially viable to build.

This capability comes with a notable shift in the human role. Mollick describes no longer feeling like a user steering a tool, but like a client commissioning a studio. The model makes hundreds of small decisions independently, with no visibility or input from the person who gave the original instruction.

Fable is not without limits. It is twice as expensive as Claude Opus, consumes tokens rapidly, and its security guardrails frequently divert requests to a less capable model. The writing style also still carries recognisable patterns from earlier Claude models.

Mollick raises a broader question: whether the growing capability of these models will inevitably mean less meaningful involvement for humans, and whether the opacity of the process is simply the price of that power.