Claude Fable 5.2 looks like a point release. It isn't. Here's what I think the model actually is, based on running Fable on real work: a different class of system that holds long tasks and keeps checking its own work; a model where the mode matters as much as the model itself; a model whose Max mode can do worse than Medium while costing five times more; and proof that you have to compare final results, not tokens.

A drive to Napa and a score of zero

My wife and I recently drove up to Napa, a wine region just north of San Francisco. On the way, I was talking with a friend, an entrepreneur. As always, the conversation turned to artificial intelligence and to the fact that I host ToTheMoon, an AI podcast from Silicon Valley. And, as always, we ran into the same thing. He said: "I use AI. We have different agents, we analyze our customers, our service runs on it."

I told him I'd give him the best advice of his life, at least when it comes to AI. Watch two ToTheMoon episodes where I break down the biggest AI trends, then rate yourself from 0 to 100 on how well you understand artificial intelligence. Right now you feel like a 60. Unless you're a deeply involved professional user, afterward you should feel like a zero.

By the way, when he said I understand AI pretty well, I told him I understand it at zero. Let me use Fable 5.2 to explain what I mean.

"So what? 5.1 became 5.2"

In November, Anthropic released Claude Fable 5.2. It seems like a small update: 5.1 became 5.2. What's the difference? Next to it sits Opus 5.5, OpenAI has its own lineups, and you can get lost in all of it.

Even Anthropic's own prompting guide notes that, in its testing, Fable 5.2 at its lowest effort level scored close to what Fable 5.1 did at higher levels. But the main thing is that Fable is a system of a different class — a different class of work altogether. To understand it, you have to start with where it came from.

A weapon released to the public

Anthropic has a model family called Mythos. It's not available to regular people: only to special companies and specially approved individuals. These models were originally built for security — cybersecurity and finding vulnerabilities. One of the best-known cases came in late spring: these models were used to analyze a large number of open-source systems and found more than 10,000 different bugs and errors.

But such a model can not only find vulnerabilities — it can create them too. That's why Mythos models are probably the most advanced weapon in the world today when it comes to cybercrime and software.

Fable is the part of that family that's available to people. Ask it to analyze your own software — a website, a homegrown CRM, a server at home — and find vulnerabilities, and it will surface a fairly large number of serious problems. It lets you close holes that human programmers could never close — not this quickly, not this far in advance, and not at this scale. I may get a pile of comments from programmers for this. But I'm a former programmer myself, I know a fair amount about building systems, and I don't believe a human was ever capable of that.

Over the summer, Fable made an incredible splash — and became the first model the US government officially blocked, after reports that it could supposedly be used to build various kinds of weapons, both cyberweapons and offline ones. In general, models at the Mythos level make it possible to create serious biological weapons too — and, at the same time, to make very serious discoveries in medicine. There are several versions of why Anthropic clashed with the US authorities, from politics to marketing, and I went through them in the full episode. The outcome: the model was unblocked and released again after several changes. Alongside Fable 5.2, Anthropic also released Mythos 5.2, still for verified organizations only.

Remember the arc: released, blocked, reopened — but with restrictions and limits. We'll come back to it.

A pool in Santa Barbara

I first tried Fable before the block. I was in Santa Barbara, my kids were swimming in the pool, and I launched Claude Code from my phone.

I was blown away by the quality of its work. For the first time, a system worked this seriously — not small stuff, but big, serious tasks. You just described the task by voice, and it knew your infrastructure, understood what you wanted, and carried the task through to the end. It held a long working cycle, which, as we know, matters a lot in development: keeping the context going, forward and deep.

Point one: it's not one session

Modern models don't work only within your single context. While it works on the main task, Fable 5.2 keeps double-checking itself and launching new sessions in parallel. You don't have just one session running — parallel sessions start up, and something is always moving.

Double-checking is one of the ideas I keep repeating from episode to episode: on every important answer, I tell the system, "check yourself." Now a strong model does it on its own as well. Anthropic's guide says that at its highest effort levels in agentic coding, Fable 5.2 used subagents to review its own work — something Fable 5.1 didn't do on the same tasks.

And in the future, the task will already be done, while the session keeps working. That's the essence. This technology is essentially already in Dots in ChatGPT — Dots, by the way, is incredible. A friend and I rented a house up north for five days to lock ourselves in and work. As soon as the Airbnb booking email hit my inbox, Dots messaged me five minutes later: "I want to point out that you have conflicts. Your daughter has a concert on Sunday, and a piano tuner is coming to your place on Monday." I never asked it to do that.

Imagine the same for your own affairs. You signed a contract, agreed to hold an event, got everything ready — and a week later the law changed. And the system tells you: "Just a reminder: the rules have changed, the price has changed." It's great when a system can keep working after you've finished a task. For now, Dots doesn't spend huge compute on this, so the monitoring is still relative.

Point two: there's the model, and there's the mode

Ask your family which version of ChatGPT they use. My wife has used it for three years and is one of its 50 million active users. Which model, which number? She doesn't know. Does she understand the modes? Unlikely. She has a $20 paid plan with Medium selected. I usually tease everyone who doesn't pick a version and tell them: "Turn it up."

But that's normal. You don't know how many megapixels your iPhone has either. What normal person would know that? Except that this has created insane confusion in the market, and it persists: Low, Medium, High, Extra High, Max, Ultra — and the modes differ from product to product.

Once again: there's the model, and there's the mode. To compare Fable and Opus and figure out which is better, you need to understand what task you're doing and which mode you're working in.

And here's the catch. On a simpler model or a lower mode, a task can take three hours — and end up only 30% solved, or not solved at all. Another mode can solve it in ten minutes, 100%. In the end, on a more powerful model whose tokens cost much more and burn faster, you can spend less actual money.

The Max paradox: when "smarter" means "worse"

Now the part that breaks people's brains. Fable's Max mode, I believe, costs three and a half times as much as Medium. And Max can do the task worse — while spending five times more money.

You might wonder: how can better reasoning be worse? That's the whole point: it can launch unnecessary agents, go in the wrong direction, and start doing things you don't need right now. It's a known fact that Fable models can lose context, stop abruptly, wander off, and launch extra agents.

Interestingly, Anthropic's own overview of Fable 5.2 says that most of the capability gain is already present at High, gains above it are uneven, and Max is for measuring the model's ceiling and for the rare task where time and tokens don't matter. At lower levels, the model is more likely to stop before the work is done; at higher ones, it delegates to subagents more readily.

For contrast, here's my most expensive story. I think it was July, right when OpenAI released Sol. I gave ChatGPT a task and walked away, and it worked on it for two and a half days. When I came back, I said: "So, did you solve it? I think you're stuck." And it replied: "I'm still working on the first item." It simply burned about 10 billion of my tokens. Fable, by the way, has never spent anything close to that for me.

Still, my advice is this. If you're not running out of tokens, use the maximum modes: in most cases, the maximum mode will give you a better result. But it's important to remember that it can do worse — many experiments won't show you that.

Don't trust comparisons. Including mine

Take the cases people share around the world — say, a system cloning a Shopify-level site. One model in one mode spent $300, another spent $900 on the same thing, or the other way around. And if you compare all these cases with each other, you won't see any correlation about what's cheaper or what's right.

If you're watching a blogger who lays out ultimate rules — which mode and which model to pick, how tokens get spent, claiming he has tested everything — that's a conditional truth. Even for him, the same task can play out differently today and the day after tomorrow: models change, you don't know which test group you landed in, which country you were in, what account you had — and whether you got lucky.

Even official benchmarks don't give a simple answer. In September, Opus 5.5 came out: on terminal-based tests — maybe not a super-important metric — it scored 66%, while Fable had 55%. And I had a big question. If Fable is a system capable of hacking anything, with a colossal level of data analysis and system building, then what is this Opus 5.5, released to everyone, that outperforms it? Or did they make Fable worse? Many people remember that Fable got worse in July. Then, about six weeks later, Fable 5.2 came out — and it wins at complex programming by a lot. And complex programming is essentially any automation. It wins by multiples.

Price per token is only half the equation, too. According to Anthropic's price list, Fable 5.2 costs $8 per million input tokens and $40 per million output tokens — 20% cheaper than Fable 5.1 and twice the price of Opus 5.5. But at one point on my $200 plan, Fable started eating all my tokens in four or five hours, and Fable 5.1 once ate my entire subscription allowance in two.

How can you even compare systems by tokens, when what you need to compare is the final result? You ask yourself: was the result achieved or not? And to figure out what works better, you have to run hundreds of different tasks, even the strangest ones.

What Fable 5.2 is for — and what it isn't

If a task is routine, more or less small, or several different answers would work for you, you probably don't need Fable. In my view, Opus 5.5 is number one today for high-volume development, building business processes, and large-scale automation. At the same time, I remain a daily ChatGPT user — for chats and basic life tasks — and a Codex user. If you can afford two $100 plans, get ChatGPT for $100 and Anthropic for $100 — you'll feel a really strong effect.

Fable is like plugging a unique brain into a task. If you need the most thoroughly verified result with the highest final efficiency, Fable is very interesting:

  • Contracts. Inheritance, a will, a divorce, the sale of a house, an apartment, or a car.
  • Finding errors in processes. What you missed in your own schedule or in tracking your own finances, which subtleties, nuances, and patterns you overlooked.
  • Finances. Financial accounting — personal and business.
  • Ads and sales. Ad analysis and sales analysis.
  • Company strategy. There are always many different options in it.
  • Complex programming and automation.

A recent example. The other day I had one fairly complex task, and Opus 5.5 was working on it. Late at night, I switched it to Fable 5.2, because I didn't fully trust Opus. There was too much money riding on it.

An honest caveat. In my opinion, Fable definitely can't fully manage advertising on its own yet. But ad analysis — absolutely. And right now is a unique time, when you can squeeze a thousand times better ad setup out of Fable. Five years ago, no system existed that could squeeze out that kind of efficiency. A time will come when this is impossible to do — I guarantee it.

What Fable 5.2 costs, and who can afford it

Let me say it right away:

  • Limits. Fable is restricted on many plans: it gets 50% of the plan. The $200 plan has no such restriction.
  • Burn rate. Fable devours tokens. There's no point using it on a $20 plan.
  • API. You're unlikely to be able to use it via the API. Some can, but be ready to set aside tens of thousands of dollars: input and output tokens cost astronomical money. It makes sense, for example, for analyzing ad campaigns if your spend is in the millions.
  • Top-ups. Buying extra tokens is inefficient. Inside a subscription, you're willing to burn tokens inefficiently; when you top up, you want to know that the $100 or $1,000 you spent needed to be spent. And an automated task can eat $1,000 on something that never needed doing — it just keeps launching useless extra agents.

If you can't pay $200 every month, you can buy the $200 plan once and try to get something done with it. For organizing your finances, I specifically recommend Fable. It's worth at least understanding that a model this powerful exists.

That's the arc I mentioned: Fable was released, blocked, reopened — and then restrictions and limits came in. Most people, using free systems or a cheap subscription, don't even come close to having access to serious models. And to do the kind of things we're doing now, over time you'll have to pay tens of thousands of dollars. That's how it should be.

Think investment, not expense

More than a year ago, Sam Altman described a very interesting case, and I think it's very viable. He said a time will come when you tell yourself, "I have $20,000 to solve this problem," hand it to AI, and it solves it. We're not there yet, because right now you pay after the fact: you can spend $20,000 and it won't even start the task, or it'll go the wrong way. Today, AI can't really plan its own spending.

But the logic itself is right. People come to me very often about starting a business. I tell them: "You shouldn't start a business. You can pay me $10,000 — I just gave you good advice, because you'll save a hundred." People aren't used to paying like that, but it's the right price. If you're putting $100,000–$200,000 into a business, and your chance of losing is about 80% from the start — and it usually is — then paying AI $5,000 to work through your problem is the right move.

People don't do this because they think in terms of expenses. They'd rather spend that $5,000 on a vacation, a hotel, a bag, or a couch than invest it in the right financial strategy for their household, a financial strategy for their business, or a marketing management system.

I often run into this exchange. "Who runs your ads?" "We do it ourselves." "But you don't know advertising." "Well, we can't hire anyone right now; our whole ad budget is $1,500." Then at least pay $300 for ad analysis by these AI systems and spend $1,200 on the ads. Actually, you don't even need to pay the $300: it's about 10% of a $200 plan.

Why this isn't just about programmers

I believe every ordinary person should now build business processes. Even my wife. Because in the near future, it will be less and less clear whom you can trust with a given task.

Right now you can still buy something without being manipulated and see something without being manipulated. But we're moving into a time when manipulation becomes far more serious than it is now. My kids are already growing up in an era of super-manipulation: there are incredibly many channels for influencing a person. And we're entering an era when artificial intelligence will truly know who you are. Everything before that was kindergarten.

I started saying this about fifteen years ago: I deeply disliked the market of deception in professional marketing. Marketers all said, "We know our customers, we've studied them, we've analyzed them." But that doesn't compare to what it means to know your customer the way artificial intelligence does now.

My rules for working with strong models

  1. Learn how to switch modes. That's number one. Don't sit on the default mode.
  2. Use Fable where you need the most thoroughly verified result. For everyday tasks, you don't need it.
  3. If you're not running out of tokens, use the maximum mode. But remember: Max can do worse than Medium and spend several times more.
  4. Work with frontier models directly, without plugins. Try a plugin only if you understand what it gives you. In Anthropic's testing, some instructions written for earlier models raised costs, caused side effects, or simply stopped working with Fable 5.2.
  5. Ask for a double-check. "Check it again for me, study the material again" — even if the model already double-checks itself.
  6. One main system and, if you have the money, one secondary. Don't spread yourself thin.
  7. Compare final results, not tokens and rankings.

So what is Claude Fable 5.2, really?

In short:

  1. It's a system of a different class of work. Fable holds a long cycle, checks itself, launches parallel sessions, and carries the task through to the end.
  2. Cost is measured per result, not per token. A more expensive model can cost you less — and its maximum mode can do worse than a simpler one.
  3. Understanding models and modes is what understanding AI means today. Most people, even on paid plans, don't know which model and which mode they're using.
  4. Right now is a unique time. Fable is already restricted by limits, and in advertising I'm sure a time will come when squeezing out this much efficiency is impossible.

The 0–100 test

Back to the drive from Napa. I asked my friend what he uses. "I have a ChatGPT subscription, I use DeepSeek." And I think he said he has Claude Code. I asked, "How much is your most expensive subscription?" "Twenty dollars." That's it — zero, right away. And he's a businessman with millions of dollars in monthly revenue. Then I heard DeepSeek, and I immediately got bored.

How do you even measure how well someone understands AI? A surgeon has parameters: how many operations, how many hours of surgery, the longest operation, where they studied, whom they worked with. A pilot has flight hours, countries flown, tough situations, aircraft types. But assessing how well a person grasps AI for their life is very hard today.

Even the people who create AI — Sam Altman, Dario Amodei — understand it very strategically, as visionaries. But many of them would find it hard to grasp the newest use cases — unless they're fanatics who build systems and create use cases themselves every day. Judging by their interviews, most of them aren't fanatics of these systems. Yet we're used to trusting: if someone talks about AI and has a hundred billion dollars, everyone assumes they understand it. And having worked as a programmer for forty years doesn't mean you understand AI. I recently read a programmer with twenty years of experience writing that he doesn't want to and won't use AI because it harms the tech world.

So your task today is to figure out what level of development and perception of AI you're at, and to keep checking against it. If you care about your professional growth, your business, your education, and your future, honestly rate yourself from 0 to 100. The gap between those who truly use AI and everyone else is already cosmic. And it will be catastrophic — and you won't even notice when it happens.

Tell me in the comments what level you see in yourself. Do you know which model you're using and why you chose it? Have you tried Fable 5.2 or Opus 5.5?


Alexander Volchek is an entrepreneur and builder, a strategic advisor to 200+ startups and scale-ups, and the host of ToTheMoon, an independent AI media outlet from Silicon Valley. Episodes, breakdowns, and company dossiers: besolid.com/tothemoon. YouTube: @ToTheMoonTech.