Ethan Mollick got early access to Claude 5 Fable, the first publicly available model in Anthropic’s new Mythos class. He gave it ambitious, vague prompts. It worked for hours at a time, spinning up its own sub-agents, making hundreds of judgment calls on its own, and delivering finished software and research projects. His takeaway: this isn’t a better chatbot. It’s a different relationship with AI entirely.
If you run a service business, that shift matters. A lot. Because the thing that’s changed isn’t just quality. It’s how little human involvement the AI needs to produce serious, complex work.
What happened
- Anthropic released Claude 5 Fable, the first Mythos-class model available to the public.
- Ethan Mollick tested it extensively on non-cybersecurity tasks (Fable’s guardrails block security work entirely).
- Fable worked autonomously for up to 12 hours on multi-page specifications, launching its own sub-agents to research, code, and verify results.
- One project: an interactive isochrone map that required researching 2,200+ flights, international rail schedules, and road speeds from academic papers. Fable launched adversarial groups of agents that checked each other’s work.
- Another project: a 9.5-hour build session that produced a sophisticated research calibration tool (called Concord) from a 19-page design document.
- Mollick’s verdict: Fable outperformed every other public model he’s used “by a considerable margin.”
The numbers
- 2,200+ specific flights researched autonomously for the isochrone map project
- 9.5 hours of continuous autonomous work on the Concord software project
- Up to 12 hours of execution on multi-page specs
- 19 pages of design documentation generated before execution began
- 2x the cost of Claude Opus per token, with very high token consumption
- Fable delegated to cheaper models (likely Claude Sonnet) to manage costs
5 things service business operators should understand about Claude 5 Fable
- The human role is shifting from “doer” to “commissioner.” Mollick describes himself not as a wizard casting spells, but as a patron funding a studio. You describe what you want, pay for it, and judge the result. The hundreds of small decisions in between? The AI handles those without asking.
- Ambition is rewarded. Mollick found that more ambitious prompts produced better results. Fable thrives on big, complex instructions rather than narrow, hand-holding ones. If you’re still writing cautious, step-by-step prompts, you’re underusing this class of model.
- It’s not perfect, and experts still matter. Mollick, as a domain expert, caught errors and omissions in Fable’s output. An independent review of its generated academic paper rated it as a “major revision” paper, meaning solid but not publishable without human refinement. The AI gets you 85% of the way. The last 15% still needs a human who knows the domain.
- Cost is real and unpredictable. Fable is twice as expensive as Opus per token and burns through tokens fast. Mollick’s description of cost in production: “a lot.” Smart delegation to cheaper sub-models helps, but budget planning for Mythos-class work is going to be tricky.
- The black box problem is growing. The more capable the model, the less visibility you have into how it made its decisions. Fable made hundreds of judgment calls on the isochrone map that Mollick never saw and couldn’t have reviewed even if he wanted to. For regulated industries, this is a real tension.
The hot take
The biggest story here isn’t that Fable is smarter. It’s that the interface between human and AI is collapsing into a single moment: the brief. Everything after that is the machine’s problem. Mollick is right that this probably isn’t a temporary artifact of bad tooling. It’s the direction. Models that can work for 10 hours unsupervised don’t need you hovering. They need you to write a really good brief and then get out of the way. That’s a fundamentally different skill than most service businesses have built. The companies that figure out how to write great briefs (clear specs, good success criteria, smart constraints) will pull ahead. The ones still trying to “manage” AI like an employee will waste tokens and time.
The Agency OS play
Here’s what to do this week, regardless of what kind of service business you run. First, pick one complex, multi-step task that currently takes your team hours of work across research, judgment calls, and assembly. Think: competitive analysis reports, client onboarding documentation, audit prep, market research summaries. Give it to Fable (or whatever Mythos-class model you can access) with a single ambitious prompt. Don’t micro-manage it. See what comes back.
Second, start building a “brief library.” If the future of AI work is commissioning rather than steering, then your most valuable asset is a collection of well-written, tested prompts that reliably produce quality output. Document what works. Note which prompts produce better results when you make them more ambitious versus more specific. This is the new operational knowledge.
Third, budget for token costs the way you budget for cloud infrastructure. Mythos-class models are expensive and unpredictable in consumption. If you’re planning to integrate long-running autonomous agents into your workflows, you need a line item for it. Start tracking cost-per-task now so you have real numbers when you scale. And don’t skip the expert review step. Fable’s output is impressive but not flawless. Build a human checkpoint at the end of every autonomous run, especially if you’re in a regulated field like law, healthcare, or finance. The AI does the heavy lifting. You do the final sign-off. That division of labor is the new normal.