Everyone Can Hire an Expert Team Now

There’s a feeling I keep having with Anthropic’s newest model, Claude Fable 5, and it took me a while to trust it. I hand it something genuinely hard, a messy business problem with ten moving parts, half of them unstated, and instead of a clarifying question it just gets to work. It holds the whole thing in its head at once. It doesn’t drop the small constraint I mentioned in passing forty minutes ago. It doesn’t forget the one instruction that mattered most. It’s curious in a way that’s hard to describe, poking at the parts of the problem I hadn’t thought through, and it checks its own work before it hands it back.

When I’m doing real business work, it feels less like using a piece of software and more like having a team from one of the Big Four sitting across the table. Not one consultant. A whole bench of them, with someone who knows tax sitting next to someone who knows operations sitting next to someone who knows the market, all of them briefed, all of them paying attention. Ethan Mollick, who teaches at Wharton and writes about this more carefully than most, landed on almost the same image the week the model came out. His old metaphor for AI was a wizard chanting a precise spell. Now, he wrote, it’s “closer to a patron” who describes what he wants and judges the result. And where a patron commissions a single artist, he said, a model like this one is “closer to a whole studio.”

I want to be honest about how much I enjoy this before I say anything careful, because the enjoyment is the tell. When a tool crosses from “useful” to “I’d rather work with this than without it,” something real has changed.

What actually changed

Not the price. Not the interface. What changed is the size of the thing you can hand off in one go.

A year ago, AI was a very fast assistant you had to supervise closely. You asked, it answered, you corrected, you asked again. Useful, but you were still doing the work; the model was just quick. What’s different now is duration and grip. The model can take a whole piece of work, hold every part of it without losing the thread, run for a long stretch on its own, and come back with something finished rather than something to correct.

There’s a group at METR that measures exactly this, and their number is the one I’d trust over any benchmark. They track the length of task a model can complete reliably on its own, and that length has been roughly doubling every several months. A year ago it was minutes. It’s now measured in hours. You don’t need to follow the research to feel the consequence: the question “what could I just hand off entirely?” gets a bigger answer every quarter, whether or not you’re paying attention. That’s the trend under the noise, and it doesn’t care what any single model is called this month.

What a model this good is actually for

Most people get this wrong in both directions, and it’s an expensive mistake either way.

If you use Fable 5 the way most people use a chatbot, to answer quick questions, draft an email, summarize a document, you’re wasting your money. The cheaper models are already excellent at that, and once this one leaves the flat subscription plans, which Anthropic has said it will, you’ll be paying premium rates for answers you could have had for a fraction of the cost. Reaching for the best model out of habit is a budget decision most people make without noticing they’re making it. I’ve written separately about the discipline of routing cheap and expensive work to the right model, and about spawning sub-agents so a strong model plans while cheaper ones execute. Those are the mechanics. This piece is about the thing underneath them: knowing what the expensive model is for in the first place.

It’s for the work you’d otherwise assign to an expert. Not a question, an assignment. Something with a real deliverable at the end, where the output has value on its own: a market analysis you’d have paid a consultant for, a strategy memo, a working prototype, a financial model built from a pile of documents. The rule I use is simple. If the result would be worth paying a skilled person to produce, hand it to the expensive model and manage it like you’d manage that person. If you’d never pay anyone to do it, you’re chatting, and you should use something cheaper.

That line is the whole filter. The casual prompter, the person typing quick questions into a box, gets almost nothing from a model like this that they couldn’t get for a tenth of the price. The person handing over a whole piece of expert work gets something that genuinely wasn’t available a year ago.

The price is the wrong number

Fable 5 costs ten dollars per million tokens of input and fifty per million out (a token is roughly a word), about twice what Claude Opus costs. It’s the most expensive model on the list, and that number is doing a lot of work in how people talk about it, most of it lazy.

Two things get left out. First, a year ago the top Claude model cost fifteen and seventy-five. Anthropic cut that by two-thirds last November. So per word, this “most expensive model ever” is actually cheaper than the best model was for most of the last two years. What went up isn’t the price of a word. It’s how many words the model spends to finish a real task, because it now thinks and works for much longer before it’s done.

Second, and this is the part that matters, cost per word is the wrong unit entirely. The question is the value of the finished result. A skilled knowledge worker in Switzerland costs the company somewhere north of a hundred francs an hour once you count everything. A serious piece of analysis is two days of that person’s time, so more than a thousand francs, before you count the weeks of hiring and the months until they’re useful. A long, hard session with a frontier model that produces the same deliverable costs, in practice, somewhere in the tens of dollars. Simon Willison, a developer who documents his AI work in unusual detail, described spending about a hundred and ten dollars of model usage in a single day and getting back what he called the equivalent of several days of work.

So the model is almost never the expensive part. The expensive part is your time writing the brief, your time checking the result, and the cost of the runs that come back wrong. Double the price of the model and you’ve added a few dollars. One failed run, a vague brief with no clear definition of done, costs you the run plus your review time plus the day you lost. Which is why the price tag is a distraction. Get the delegation right and the model cost rounds to zero next to the value of the outcome. Get it wrong and it wouldn’t matter if the model were free.

The honest limits

I’d be selling you something if I stopped there, so here’s the other side.

It’s slow. All that thinking takes time, and for a quick task it can feel like a crawl. That’s a real reason to keep it for the work that deserves it and use faster models for everything else.

It refuses more than it should. Anthropic wrapped this model in safety filters aimed at serious misuse, and in the early weeks they’ve been clumsy, blocking legitimate work in biology and security that happens to sit near a tripwire. If your work lives in one of those areas, test before you commit.

And there’s a data rule that catches serious users off guard. Fable 5 requires that your prompts be retained for thirty days, with no zero-retention option. For a lot of regulated work, healthcare records, legal matters, sensitive financial data, that alone can rule it out. Microsoft reportedly restricted its own staff from using the model within days of launch over exactly this. Check what you’re allowed to send before you send it.

Then there’s the thing that convinced me not to build anything important on a single model. Three days after Fable 5 launched, the US government ordered Anthropic to suspend it over an export-control concern, and because the order was broad and there was no clean way to comply narrowly, the company switched it off for everyone, worldwide, by its own account. It was gone for nineteen days, from the twelfth of June to the first of July, when the order was lifted and it came back with new safeguards. I’m not here to argue the politics. The operational lesson is what stuck with me. A frontier model can vanish overnight for reasons that have nothing to do with you. If your whole operation runs on one, you’re dark until it returns. If the frontier model is the brain you plan and review with, while the day-to-day work runs on cheaper, widely available models, you bend instead of breaking. The capability that matters isn’t any one model. It’s your way of working, and that has to survive losing your favorite tool.

Who this leaves behind

Now the part that’s my opinion, and I’ll flag it as opinion.

Everyone can rent the expert team now. That’s the genuinely new thing. The consulting bench, the studio, the specialist for every corner of a problem, available to a company of five the same as a company of five thousand. So the advantage stops being access to the capability. Everyone has access. The advantage becomes knowing what to do with it.

And that advantage compounds in a way you can’t buy back later. Every time you delegate a piece of real work well, you leave something behind: a brief you can reuse, a chunk of context the models now understand about your business, a check that catches a class of mistake before it ships. The next delegation is faster and more reliable because of the last one. It accretes into the organization, into how the place works. That’s not a product you can purchase in a hurry when you finally notice you’re behind. It’s accumulated judgment, built one real task at a time, and money can’t compress it. Throwing consultants at it doesn’t work either, because the people who actually understand this are busy building their own things.

You can already see the shape of it. The startups people talk about for their valuations are quietly more interesting for their headcounts: a handful of people producing what used to take hundreds. From what I can see, the revenue per employee at the sharpest of them runs many times what a normal software company gets. So here’s my forecast, and it’s a forecast, not a fact. Within a few years, the companies that learn to direct this well will run at a large multiple of the output per person of the ones that didn’t. Maybe I’m being too dramatic, but think it through. If one person can reliably run several delegated pieces of expert work at once, then ten people out-producing a hundred stops sounding like a slogan and starts sounding like arithmetic. And by the time the slower company feels the pressure in its numbers, the gap won’t be closable, because the company that’s ahead won’t be ahead by something you can order. It’ll be ahead by two years of compounded practice.

I owe you the other side, because it’s stronger than the boosters admit. Big incumbents have real moats, distribution and data and trust and regulation, and disruption always takes longer than the people predicting it want it to. Klarna announced its AI had done the work of seven hundred support agents, then walked it back a year later when the quality didn’t hold, and rehired people. Plenty of real work still doesn’t compress cleanly, the parts full of clients and politics and things that happen in the physical world. And there’s a ceiling I keep hitting myself: the work can scale faster than my ability to check it. Output grows with the models. Review still grows with people. Until you can delegate the checking too, that’s a hard limit.

None of that changes the direction for me. It changes the timeline, and it should make you humble about the exact year. But the direction is not in much doubt.

Where to start

Don’t rush out and put everything on the expensive model. Do one thing this week instead. Take a real piece of work, something whose result would actually matter, and write the brief you’d hand a competent expert: what you want, what the constraints are, what “done” looks like. Then give it to a strong model and read what comes back against what you’d have gotten from a quick question in a chat box. Pay attention to how much of the value was in the brief, not the model. That’s the whole skill, and it costs you an afternoon.

For me, after the drama and the shutdown and the bill, the answer to “is it worth it” turned out to be yes, but narrowly. I use Fable 5 to think and to plan and to check, for the hard problems where a real thinking partner earns its keep, and I route the rest to cheaper models it briefs. The pleasure of working with it is real, and I’m not going to pretend otherwise. But the pleasure isn’t the point. The point is that the expert team is available to everyone now, and the only thing that’s scarce anymore is knowing how to run it.