Claude 5 Fable a passé des heures à étudier 2 200 vols pour construire une carte isochrone à partir d'un prompt vague

I gave a single vague prompt to Claude 5 Fable—the first Mythos-class AI model released to the public—and it spent hours launching sub-agents, researching 2,200 specific flights, and building a fully functional isochrone map from scratch while I sipped coffee. No previous model came close on this task.

Ethan Mollick of Wharton got early access to Fable through Claude Code and documented what happens when you give an AI an ambitious, open-ended instruction. The results are both delightful and unnerving. Delightful because it actually works. Unnerving because the human contribution shrinks to a few sentences of feedback.

Autonomous Sub-Agents Researching Real Travel Data

Mollick prompted Fable to "build a fully researched and beautiful isochronic map" based on real data, covering airports, trains, walking, and driving. Fable responded by launching multiple cheaper Claude Sonnet agents to conduct parallel research: it retrieved over 2,200 specific flights, rail schedules from the TGV to the Shinkansen, and road speeds per country from multiple academic papers. While those agents ran, it started coding.

Then it launched more agents and tests to verify its own code, taking notes about its progress. The resulting map looked like the original 1881 isochrone map from London Mollick referenced. But it wasn't perfect—remote locations like Greenland fell back to estimates. A second instruction to "actually get travel times to remote airports and locations" triggered an adversarial workflow of research agents cross-checking each other. Fable figured out how often ships sail to Pitcairn Island and how to get to Grise Fjord from Ottawa.

Minimal Human Oversight, Maximum Autonomous Execution

The unnerving part: Mollick's role was extreme vagueness. "I just asked for something and it happened." He gave a single ambitious instruction, then a couple of minor feedback lines, and the AI figured out the rest—including research, math, visual development, taste, judgement, and complex coding. Fable worked up to a dozen hours executing on multi-page specifications.

Other tests confirmed the pattern. Fable created a 10-page epic rhyming poem about a haircut where every word starts with the letter 's'. It produced games like a Balatro-style coin-flip game, a self-aware snake game, and an art game inspired by the Duino Elegies—all generated from math alone, no images, since Claude cannot generate images.

What This Means for How We Work with AI

Mollick's post isn't about benchmarks; it's about a qualitative shift in the human-AI relationship. Previous models couldn't handle the multi-step judgment calls required by the isochrone map task. Fable treats a vague prompt as a project spec and deploys its own orchestration to execute it. The model essentially becomes a manager of sub-agents, a coder, a QA tester, and a researcher rolled into one.

Fable's guardrails prevent cybersecurity use, so the security implications of Mythos-class models are separate. But for creative, research-heavy, and software development tasks, the boundary between tool and collaborator just got a lot blurrier. Expect more posts like Mollick's—not because the demo is flashy, but because the shape of work itself just shifted again.

Source: What it feels like to work with Mythos
Domain: oneusefulthing.org