Anthropic's smartest model is now included for Max subscribers. No meter, no weekend deadline, just usage limits that reset like everything else. Which changes the question. It's no longer "what do I run before the price kicks in." It's "what actually deserves the smartest model ever shipped?"
Here's the wrong answer: more autocomplete. If you point Fable at boilerplate a cheaper model writes fine, you're using a chief of staff to fetch coffee.
Here's what I'd do instead.
I confess I did this wrong for a while. My prompts were polite little work orders. "Review this code." "Summarize this doc." "Check this for bugs." The output was fine. Fine is the problem. A reviewer with no skin in the game gives you reviewer answers.
Then I started handing the model an identity with something to lose, and the output changed completely.
There's a technical reason this works on Fable specifically. Anthropic's own prompting guidance says to give it goals, not checklists. Fable is built to run long, handle ambiguity, and decide what matters on its own. Step-by-step instructions that helped older models actually drag it down. An ownership stance is the purest goal-shaped prompt there is: here's who you are, here's what's at stake, go.
A consultant gets tasks. An owner gets consequences. Prompt for consequences.
Below are the 10 stances I use. Each one pushes the model into a different corner of ownership, so you get variety instead of ten flavors of "looks good to me."
"Audit my codebase and database like you're the new CTO on day one, not a consultant billing hours. What is everyone too close to see? What would you kill that people are afraid to touch? End with one contrarian 90-day bet, and the strongest argument against your own bet."
That last line matters. Fable is good at self-verification when you ask for it. Making it argue against its own bet is the difference between a hot take and a decision you can act on.
"You just bought 20 percent of this company and you only make money if it doubles in two years. Read everything. What do you force us to stop doing at the first board meeting? What's the one number you'd tie every paycheck to, and what would prove in six months that you picked the wrong number?"
Ownership with a deadline attached. Two-year money has no patience for a five-year story.
"You were my cofounder for two years and just left. Write the brutally honest note you'd never say to my face: what's broken, what I'm avoiding, and the one thing you hope I finally do without you."
This one stings. Run it anyway.
"You're running technical due diligence before a $10M acquisition of this company. What would make you walk away? What's the buried problem that isn't obvious yet? For each finding, say what evidence would prove you wrong."
Note the framing. Business risk, not "find vulnerabilities." Fable's safety margins are wider than any prior model, and security-research phrasing is the fastest way to get a refusal. Due diligence framing gets you the same rigor without tripping the wire.
"Rebuild this from scratch knowing everything you now know about it. What's the first architectural decision you'd change, and why didn't we make it the first time? Separate 'we couldn't have known' from 'we didn't want to know.'"
"Cross-reference the codebase with our support tickets and customer complaints. What root cause has nobody connected yet? Show the ticket-to-code trail so I can check your work."
The trail requirement keeps it honest. If it can't show the connection, it doesn't get to claim the connection.
"Find the most over-engineered thing in this codebase and the most under-engineered thing. Tell me which one is actually costing us more, in hours or dollars, and what number you'd track to prove it."
"Read the codebase and the org structure together. Where does the code reveal a team or communication problem that isn't visible from the outside?"
Your architecture is a fossil record of your org chart. The model reads fossils fine.
"You have to run this entire company alone starting tomorrow. What do you automate first, delete first, and double down on first? Assume you only get 20 hours a week."
The 20-hour constraint is doing the work here. Unlimited-time answers are wish lists. Constrained answers are strategy.
"This company is still alive in five years. Which decision made this quarter do people point back to as the mistake? Fix it now, while it's cheap, and tell me what it costs to fix today versus in a year."
Set effort to high. Anthropic recommends high as the default and xhigh for the most capability-sensitive work. Audits like these are exactly that work.
Don't ask it to show its reasoning. A standing "explain your reasoning" line can trigger a refusal, and in some setups Fable silently hands your task to Opus 4.8 and burns your usage doing it. (If that's been happening to you: Settings, then Capabilities.) Ask for conclusions and evidence, not the inner monologue.
Spend your limits where they compound. Stance audits are token-hungry. They read your whole codebase, your tickets, your docs. That's exactly why they're worth the budget: one good audit changes what you build for a quarter, while a hundred autocompletes change nothing. Start your usage window with the big questions, not the small tasks.
The models got smart enough to be owners. Most people are still prompting them like interns.
