Anthropic Releases Claude Fable 5.2 and Claude Mythos 5.2…

New crowdsourced testing promotional copy. New crowdsourced testing promotional copy end. Add App-exclusive traffic redirect

(Source: AI Information Gap)

Just now, Claude Fable 5.2 and Claude Mythos 5.2 were officially released. On almost every test, the two new models score a notch higher than Claude Opus 5.5, and API pricing is also 20% lower than the previous generation.

Claude Pro users will also be able to try Fable over the next week, while Claude Max and Ultra users can claim 1 to 3 Fable quota resets based on subscription length.

Just now, Anthropic officially released Claude Fable 5.2 and Claude Mythos 5.2.

1,000 Claude Fable 5.2 agents took only 30 hours to fully formalize the proof of Fermat's Last Theorem. With the same configuration but the model switched to Claude Opus 5.5, it ran for 18 hours and completed only 26%, then barely made any further progress.

That's how strong it is.

In Anthropic's capability tests, Claude Fable 5.2 scores a notch higher than Claude Opus 5.5 on almost every item. Compared with the previous generation Claude Fable 5.1, its output speed is 30% faster and its price is 20% lower.

Key point: Claude Pro users can also use Fable this time.

So-called formalization means translating a mathematical proof into a language that a computer can check step by step for correctness.

The whole process involved almost no human intervention; how to divide and coordinate the work was decided by the agents themselves. They also did not use mathematical reference materials—they had only a list of intermediate theorems for scoring.

The last time Anthropic formalized Fermat's Last Theorem, it used an unreleased research version of Claude, took 11 days, relied on a specially built framework, and had guidance from a researcher.

Another official demo is even more outrageous.

An Anthropic engineer asked Claude Fable 5.2 to develop an operating system starting from an empty folder.

It orchestrated more than 1,000 sub-agents, ran continuously for 8 days, and finally delivered a system that can boot on a laptop, complete with drivers, file storage, windows, a login screen, an installer, and Claude built in.

Now, that's no longer necessary.

From the time you see this article until November 17 Beijing time, Claude Pro and Team standard seats can use half of their weekly quota for Fable at no extra charge, and also receive 1 Fable quota reset.

Open Claude, and you will most likely see this pop-up.

Claude Max, Ultra, and Team premium seats receive Fable quota resets based on subscription length: 1 for less than half a year, 2 for half a year to one year, and 3 for more than one year, valid until March 9 next year Beijing time.

Key point: this is a reset card exclusive to the Fable model, and it can only restore the portion of the weekly quota used by Fable (for Max, the full 5-hour quota is restored).

The reset card arrives automatically and looks like this.

To use it, open "Usage" in Claude settings and click "Reset for free."

As expected, Claude Fable 5.2 is still strongest at programming.

The programming benchmark FrontierCode 2.0 examines whether AI-modified code can be merged by maintainers of open-source projects. With reasoning intensity set to Xhigh, Claude Fable 5.2 scored 49.1%, the highest overall.

Even better, at the High setting it scores 45.3%, almost on par with Claude Opus 5.5 maxed out at Max (45.5%), while the cost per task is only about half.

Now look at this kill line for top models. The previous generation Claude Fable 5.1 has already been ruthlessly knocked out.

The horizontal axis is cost, and the vertical axis is score.

The biggest improvement is in robot control.

On RoboDojo, a test simulating a dual-arm robot, Claude Fable 5.2 scored 64.9%, while the previous generation Claude Fable 5.1 scored only 27.6%—more than double.

Beyond simulators, Claude Fable 5.2 can already control real robots: unscrewing bottle caps, solving a Rubik's cube, and drawing.

On the model leaderboard of third-party evaluator Vals AI, Claude Fable 5.2 currently ranks first with a comprehensive score of 71.31%.

Claude Fable 5.2's "sensitive skin" has improved considerably this time.

"Sensitive skin" is an old problem of the Fable model: the safety classifier often blocks normal requests as high-risk. Compared with Claude Fable 5.1, 5.2 has about 55% fewer false blocks.

When your request really is blocked, another model will answer instead. Cybersecurity questions go to Claude Opus 4.8, and biology questions go to Claude Opus 5.

Claude Mythos 5.2 is Claude Fable 5.2 without this layer of safety protection, so it is only open to approved cybersecurity and biomedical research institutions.

One commenter's joke hit the nail on the head: "Fable and Mythos are the same model, except one of them is licensed to work."

Leaks about Claude Fable 5.2 began as early as early October.

On October 3, a netizen asked Claude to "generate an animation about a point," and it delivered a Pixar-quality short film, with a few hand-drawn words at the end reading Claude Fable 5.5; we followed up immediately.

At the end of October, a mysterious dark horse, Obsidian 3, emerged from the raw data of FrontierCode 2.0, scoring higher than Claude Opus 5.5, but it was wiped clean that same day.

In early November, Claude Mythos 5.5 appeared again in the source code of Anthropic's official website.

Just before the official announcement, some netizens also found the prices of Claude Fable 5.2 and Claude Mythos 5.2 in advance in Amazon Web Services' official pricing documents.

You'll need to swipe for a while on this one.

The pricing later announced by Anthropic was completely consistent with the netizens' leak: $8 per million input tokens and $40 for output. The previous generation was $10 and $50.

Stronger, and cheaper.

In earlier leaks, the new model's name was always 5.5, but at official release it became 5.2.

In version number, Claude Fable 5.2 < Claude Opus 5.5; but in capability, Claude Fable 5.2 ≫ Claude Opus 5.5.

For everyday coding, research, and document writing, the more cost-effective Claude Opus 5.5 is recommended first; when facing the toughest problems, bring out Fable.

Claude Pro users can try Fable without paying extra, until November 17 Beijing time. Get on board quickly. Statement: Exclusive article from Sina.com, unauthorized reproduction prohibited. tech_web_article_v2015_block_bottom Non-targeted 300*250 button 17/09 wenjing begin

Non-targeted 300*250 button end