Claude Fable's marathon autonomy impresses — and turns AI into a black box

Ethan Mollick got early access to Claude 5 Fable, the first publicly released Mythos-class model, and reports a substantial capability jump over every prior model he has tested. Fable sustained autonomous work sessions lasting up to twelve hours against multi-page specifications, producing results ranging from a sophisticated AI-generated social science paper to playable games built entirely from one vague prompt — with all art generated mathematically, since the model cannot produce images. Notably, Anthropic’s guardrails block the model from cybersecurity work entirely, so Mollick’s testing deliberately covered everything else.

Two projects illustrate the shift. Asked to build a researched isochrone map — a task no previous model handled even passably — Fable spawned fleets of cheaper sub-agents to gather over 2,200 flight records, rail timetables, and per-country road speeds while it wrote and tested code in parallel. When told to replace estimated travel times for remote locations, it ran adversarial agent groups that researched and cross-checked each other, working out ferry schedules to Pitcairn Island in the process. A second project, a nine-and-a-half-hour build of a research tool called Concord for calibrating human and AI judgment in messy datasets, delivered software Mollick says researchers have wanted for years but that was never commercially viable to build.

The unnerving part, Mollick argues, is not the output quality but the collapse of human involvement. He gave ambitious instructions and light feedback; the model made hundreds of judgment calls invisibly, with reasoning too voluminous to audit. The work was imperfect — an expert eye still caught errors — but the human role shrank to specification and spot-checking, making frontier AI both more capable and more opaque at once.