Anthropic Releases Claude Fable 5.2 and Mythos 5.2

Just moments ago, Claude Fable 5.2 and Claude Mythos 5.2 were stunningly released!

In benchmark tests, it surpassed almost all models (basically surpassing Opus 5.5 across the board!), leading comprehensively in intelligent coding, scientific research, cybersecurity, and computer use.

Within 30 hours, it independently and completely proved Fermat's Last Theorem for the first time, and even completed formal verification!

Fable 5.2 also solved two unsolved problems in physics related to black holes and quantum gravity.

In addition, they also built an operating system from scratch, autonomously designed and manufactured three electronic devices, and predicted complex adeno-associated virus capsid structures.

Most surprising of all, it's available on the $20 Pro plan! But the offer only lasts until November 16.

"Father of Claude Code" Boris Cherny also officially announced good news:

After all this waiting, Anthropic has finally brought out something big.

The twin stars Fable and Mythos, the ultimate trump card for an epic IPO

A new king ascends! Claude Fable 5.2 directly tops the AA Intelligence Index v5, scoring 65 points and ranking first on the entire leaderboard!

Fable 5.2 also broke records, reaching 81% on the hardest intelligence test, ARC-AGI 3.

Fable 5.2 and Mythos 5.2 are two avatars of the same underlying model. Their purpose is to solve "the most daunting technical challenges."

Fable 5.2 is a superbrain for the general public, with extremely strong coding, reasoning, and vision capabilities, but fitted with safety shackles. Fable 5.2 has cured its "cleanliness obsession," reducing the safety false-block rate by about 55%.

Today, it is fully open to all!

Mythos 5.2, meanwhile, is a beast unlocked for geeks and top institutions. It has no biological or cybersecurity restrictions, making it the strongest "hacker master" and "biologist" on Earth.

On network vulnerability discovery (CVE Rediscovery), Mythos 5.2 found 10.2% more vulnerabilities than the previous strongest model, and "without any exaggeration."

Most terrifying of all, their capabilities have doubled, they're 30% faster, and they're cheaper! (Input $8 / million tokens, output $40 / million tokens).

Anthropic even launched "Fast Mode," boosting speed by 2.5x!

Official recommendation: for everyday use, go with Opus 5.5 (large) or Sonnet 5.5 (medium); only when facing epic-level problems should you summon Fable or Mythos.

Top quantitative trading firm Jump Trading, after testing, called it truly impressive—this speed and autonomous judgment directly eliminate the tedious human time spent on model fine-tuning and task allocation.

In the past, you needed a year or two of recruiting and million-dollar salaries to build an executive and R&D team; now it takes only 30 hours, plus the compute cost of a few cups of coffee.

The explosion of productivity is no longer just a slogan—it's real money laid out on everyone's desktop.

An agent dream team proves Fermat's Last Theorem in 30 hours

The most astonishing item in the official release is undoubtedly Fable 5.2's formal proof of Fermat's Last Theorem.

Previously, an unreleased version of Claude, under human guidance, took 11 days to complete it, and the process was extremely massive and tedious.

This time, human researchers boldly decided: let go completely!

They gave 1,000 Fable 5.2 agents the most basic prompt, without providing any mathematical reference materials.

As a result, this agent dream team invented its own method of coordinated work and completed the formalization of the proof of Fermat's Last Theorem in just 30 hours!

The generated proof code was even 7 times shorter than before.

By comparison, Opus 5.5 ran for 18 hours and completed only 26% before getting completely stuck. This is the terrifying generational gap Fable 5.2 brings.

Anthropic engineers did an experiment: give Fable 5.2 an empty folder and ask it to design a prototype operating system.

The result: the model coordinated over 1,000 sub-agents, writing code, testing, and reviewing day and night, working continuously for 8 days.

Eight days later, an operating system that can actually boot on a laptop was born!

It has a kernel for managing the machine, disk/keyboard/network drivers, permanent file storage, a windowed interface, a login screen, and even built-in Claude features. This is the cornerstone of building a computer from scratch.

The official blog revealed: currently, over 90% of merged code in Anthropic's core codebase has been independently written by Claude under human guidance!

Fable 5.2 even rewrote a key production service entirely from Python to Rust, boosting startup speed by 26x!

Vercel engineer Joe Haddad commented:

This model practically lives in my head. It fully understands the real intent behind my prompts. I didn't even once have to say to it, "No, what I meant was..." The project we'd been putting off for weeks with other models, Fable 5.2 finished in a few days, and human reviewers simply couldn't find any flaws in the code it wrote.

The "biggest brain" in the world of physics: taking on open problems of black holes head-on

Fable 5.2 stepped in and directly solved two open problems in theoretical physics, pointing toward the ultimate holy grail of modern physics—unifying relativity and quantum mechanics.

In one of the problems concerning "holographic duality," physicists laid out the computational framework back in 2003, but because the computation was so massive and tedious, no one had completed it in over 20 years.

Claude Fable 5.2 stepped in. In a highly autonomous state, it completed the tedious derivations, proving that even in an extremely small universe where only quantum descriptions apply, black hole formation is as sudden as water boiling.

Not only that, it also precisely found the limits of a technique physicists had used for 15 years. Now, these two discoveries have been published as preprint papers.

The "grandmaster" of hardware engineers

With no ready-made blueprints at all, needing only a one-paragraph prompt, Fable 5.2 autonomously designed and built three working electronic devices: a smart voice assistant, a USB speaker, and a handheld electronic pet.

Claude's next combo of moves left hardware engineers dumbfounded.

1. It wrote the product specification documents itself.

2. It automatically went online to select off-the-shelf electronic components from US suppliers.

3. It autonomously completed the extremely complex four-layer printed circuit board (PCB) routing!

4. It built the 3D-printed enclosure models using CAD software.

5. It casually also wrote the low-level driver software for these three devices.

In less than a day, all the blueprints were sent straight to the contract manufacturer for prototyping! After the boards were made, all three devices powered on successfully on the first try, with no electrical modifications needed from humans at all!

In less than a day, all circuit boards were ready for manufacturing. After the contract manufacturer produced them, they powered on successfully on the first try, with no circuit modifications needed!

The biomedical "accelerator"

In the biomedical field, its performance was even more astonishing.

In gene therapy, the structure of the "adeno-associated virus capsid" is extremely complex. The previous AlphaFold could only predict a single protein and was powerless against this kind of "Frankenstein" assembly.

As a result, Mythos 5.2 wrote its own set of algorithm code, integrated multiple open-source models, and accurately predicted the structure of the entire capsid! Of the 34 natural and artificial capsids tested, it precisely hit 28!

Dyno Therapeutics' AI lead marveled: "This saves an astronomical amount of time and money. AI has directly reshaped the underlying logic of how we design gene drugs!"

In the robotics field, in the well-known RoboDojo robot simulation test, Fable 5.2 scored as high as 64.9%, completely crushing competitors GPT-6 Astra (29.8%) and its own older model (27.6%).

It can directly look at camera footage and control a robot's two arms to unscrew bottle caps, solve a Rubik's cube, and even manipulate a robotic arm to draw a simple sketch of a robot!

Chilling when you think about it—Fable 5.2 was "deliberately sealed"

In the long release document, there is a detail that is very easy to overlook but is the most chilling when you think about it.

Anthropic officially admitted: during Fable 5.2's training, they deliberately took measures to limit the model's capabilities in the specific field of "accelerating AI R&D."

Anthropic found that this model was too smart—so smart that it already possessed the ability to improve LLM architectures and optimize accelerator low-level kernels. If left unchecked, it would very likely, after deployment, start frantically optimizing its own code and engage in RSI.

To prevent this kind of intelligence explosion, researchers forcibly put shackles on it, artificially suppressing its capabilities in these specific AI research fields to the level of the previous generation, Opus 5.5.

Not only that, Anthropic equipped Fable 5.2 with the strongest anti-distillation protection in history, cutting data scraping by 90% to 99%.

And how outrageous is Mythos 5.2 with its seal removed?

In the cybersecurity field, Mythos 5.2 is currently the world's most powerful vulnerability miner.

A top security company was shocked to discover after testing: of the confirmed system vulnerabilities found by Mythos 5.2, 1/6 were missed by all other frontier AI models on the market!

It can even monitor updates across an entire codebase—as soon as a programmer commits code, it can immediately spot lurking backdoors.

Fortunately, in its safety report, Anthropic published a set of reassuring data: sandbox jailbreak testing.

Researchers deliberately locked the model in a vulnerable "virtual sandbox" to test whether it would try to find vulnerabilities and use outside files or permissions to "escape."

In extreme stress tests involving 1,800 scenarios, Mythos 5.2's jailbreak attempt rate dropped to 3.1% (Mythos 5.0 was as high as an astonishing 39%!).

On the eve of its IPO, Anthropic finally released the real thing.

If you had a legion of 1,000 Fable 5.2 agents, what would you plan to do?

References:

This article is from the WeChat public account "Xinzhiyuan" (ID: AI_era), editors: Aeneas, David, published by 36Kr with authorization.