Claude Fable 5.2: The Most Powerful AI, and You'll Have to Pay Dearly for It?

Alexander suggested to an entrepreneur friend who thinks he understands artificial intelligence that he watch two episodes on trends and honestly give himself a score from zero to a hundred. After them, he believes, most people should feel that they know AI at zero — unless the person is a professional user deeply involved in the subject.

Fable comes from the Mythos family of models. They are not available to ordinary people and are given to special companies and people, because they were originally built for cybersecurity and finding vulnerabilities; in the spring, according to Alexander, they were used to find more than ten thousand bugs in open-source systems.

A model that finds vulnerabilities can also create them. That is why Alexander calls the Mythos models probably the most high-class weapon in the world in terms of cybercrime and software, while Fable, in his view, closes holes in a way human programmers never could.

Fable became the first model officially blocked by the American government: the authorities considered that it could threaten national security and that it could supposedly be used to create cyberweapons and offline weapons. Such models, according to Alexander, make it possible to create biological weapons too — and at the same time to make serious discoveries in medicine.

Polina has been using ChatGPT for three years, and it has accumulated a huge amount of data about her. Alexander suggests thinking about how well a person understands what is stored inside such systems and what is collected on their devices — and remembering that the rules on what one may ask Grok depend on Elon Musk's strategy and can change.

According to one version, Anthropic became unfriendly to the government after the capture of President Maduro, in which its model was used: the company's leadership objected, and the White House threatened to exclude it from all systems. Alexander attributes the call threatening prison to the head of a leading AI company, which the US president described, to the head of Anthropic.

Alexander comments on the scandal around Jensen Huang with irony, but something else matters more to him: all the world's super-cool infrastructure is built on NVIDIA chips, and that is why catching up with America is practically impossible — even China, with its own chips, tries to buy NVIDIA.

One of Alexander's theories is that the story of the block could have been marketing for Anthropic. The four market leaders — Google, OpenAI, Anthropic and xAI — are incredibly closely connected, while xAI, in his words, is an obvious world leader in AI, and arrangements in the market are possible.

Before the block, Alexander launched Fable in Claude Code from his mobile phone in Santa Barbara — and for the first time saw a system that doesn't do small stuff but, from a voice description and knowing the infrastructure, carries a large task through to the end and holds a long work cycle.

Modern models, especially Fable 5.2, re-check themselves in parallel while working and launch new sessions; a task may be done while the session keeps working. The Dots feature in ChatGPT is built on this: after a house booking it reminded Alexander on its own about his daughter's concert and the piano tuner's visit.

After the block, Fable was brought back with restrictions and limits, and on a cheap Anthropic plan such models can't be used. But sometimes, Alexander believes, it is worth paying an extra hundred, two hundred or three hundred dollars: this is a system of a different class of work.

Against Fable, OpenAI released Sol with a long context window and Astra, while Anthropic released Opus 5.5. Alexander considers it today's top model for mass development, business processes and automation, while remaining a daily user of ChatGPT and Codex himself.

So many models came out over three years that people stopped understanding them: Polina, one of ChatGPT's fifty million active users, doesn't know its version or the difference between Medium, Pro and High. Alexander compares it to megapixels in an iPhone — the confusion in the market is insane, and it persists.

There is the model, and there is the mode. A task in a simple mode may take three hours and be thirty percent solved, while in another it takes ten minutes and is fully solved. That is why, according to Alexander, you can end up spending less money on a more powerful model with expensive tokens.

If you compare the published cases — which model in which mode spent how much on a copy of Shopify or another store — no correlation as to what is more cost-effective and more correct is visible. Bloggers' rules on choosing a mode are a conditional truth: models change, and the result depends on the country, the account and the test group.

In July, when Sol came out, Alexander set ChatGPT a task and left, and two and a half days later it replied that it was only working on the first item — about ten billion tokens went that way. Fable has never spent that much for him, although it too is known for losing context and launching extra agents.

Fable in Max mode, which costs about three and a half times more than Medium, can do a task worse and spend five times more money: deeper reasoning launches extra agents and goes the wrong way. But if your tokens allow, Alexander still advises the maximum modes — in most cases they give a better result.

Everyone will have to build business processes, because it will soon be less and less clear whom to trust with a task: manipulation will become far more serious. We are entering an era when artificial intelligence will truly know who you are — compared with that, marketing that "studied its clients" was kindergarten.

Fable is worth giving tasks that need a maximally verified result: drafting a contract on inheritance, divorce or a sale, finding an error in a process, seeing what was missed in a schedule or finances, analyzing advertising, sales and company strategy. For everyday small tasks there is probably no need to bring it in.

On many plans Fable is limited to half the quota, on the two-hundred-dollar one it is not, and it burns through tokens very quickly, so there is no point using it on a twenty-dollar plan. Via the API it costs astronomical money and is justified perhaps only for analyzing advertising with spend in the millions of dollars.

Sam Altman described a time when you will be able to allocate twenty thousand dollars to artificial intelligence for a task, and it will solve it. We are not there yet: the model cannot plan its costs in advance, and money is spent after the fact — sometimes with no result.

If you invest a hundred to two hundred thousand dollars in a business with a probability of losing of around eighty percent, paying five thousand to artificial intelligence to work through the task is, according to Alexander, the right move. But people think in terms of expenses and would rather buy a bag or a sofa.

At some point Fable was eating up Alexander's two-hundred-dollar subscription in four or five hours. Buying more tokens is inefficient: when buying more you want to know that a hundred or a thousand dollars were not spent on unneeded agents or on what nobody asked for.

A surgeon and a pilot have parameters for assessment — operations, flight hours, aircraft types — but there is no such assessment for artificial intelligence. Even AI creators like Sam Altman and Dario Amodei understand it strategically, but many of them, unless they are fanatics, find it hard to grasp new applied use cases.

Forty years of programming or a hundred billion dollars do not mean an understanding of artificial intelligence. The task of everyone who cares about professional development, business and education is to determine what level they are at and keep checking against it.