Anthropic Releases Claude Fable 5.2 and Mythos 5.2
Anthropic Releases Claude Fable 5.2 and Mythos 5.2
Anthropic is giving the AI industry another reason to pay attention.
Just days after OpenAI introduced GPT-6 Cyber, Anthropic has released two upgraded AI models: Claude Fable 5.2 and Claude Mythos 5.2.
These are not just routine updates with a few behind-the-scenes improvements.
The company says the new models are faster, better at complex coding tasks, and capable of handling technical challenges that previously required significant human effort. Its research team also reported some remarkable experiments, including designing working electronic devices and using more than 1,000 AI agents to formalize a famous mathematical theorem.
But there is another interesting detail: Anthropic is offering the two models under different access rules. One is available to a much wider audience, while the other is reserved for organizations that pass additional checks.
The announcement, made on November 9, 2026, also puts Anthropic in a stronger position against competitors such as OpenAI and Google, particularly on price and software development.
Meet Claude Fable 5.2 and Claude Mythos 5.2
The first thing to understand is that Fable 5.2 and Mythos 5.2 are built on the same underlying model. The difference is largely about access and safety restrictions.
Claude Fable 5.2 is the version designed for general users and businesses. It can handle demanding coding projects, technical work, reasoning, and other complex tasks while maintaining safeguards around sensitive activities.
Claude Mythos 5.2, on the other hand, is designed for more specialized work. It provides access to advanced cybersecurity and research capabilities, but Anthropic restricts it to vetted organizations through its Cyber Verification and Life Sciences programmes. Think of it as one powerful engine with two different access levels.
Anthropic wants developers and businesses to benefit from the improved model, but it does not want every user to have unrestricted access to capabilities that could create serious security risks. This approach reflects a growing trend among AI companies: making powerful models available while controlling access to their most sensitive capabilities.
The New Models Are Up to 30% Faster
Speed is one of the headline improvements.
Anthropic says Fable 5.2 and Mythos 5.2 can generate responses up to 30% faster than their respective 5.1 versions. That matters because developers and businesses increasingly use AI for tasks that involve large amounts of code, long documents, and complicated instructions.
A model that can produce useful results faster can reduce waiting time and help teams complete more work. However, the 30% figure is an upper limit reported by Anthropic, not a guarantee that every task will become 30% faster. Actual performance will depend on the type of work being done.
The company has also introduced an optional fast mode that can increase generation speed by up to 2.5 times. This is intended for users who need faster responses and are willing to pay more for the additional speed.
Anthropic Says Its AI Helped Prove a Famous Mathematical Theorem
One of the most interesting claims from the launch involves mathematics.
During pre-release testing, Anthropic says its researchers coordinated more than 1,000 Fable 5.2 agents to formalize a proof of Fermat's Last Theorem within approximately 30 hours. This is a famous mathematical problem that remained unsolved for more than 350 years before mathematician Andrew Wiles published a proof in the 1990s.
So, what does formalizing the proof mean? It means translating mathematical reasoning into a precise, machine-checkable format that a proof assistant can verify.
The important distinction is that the AI was not independently discovering Fermat's Last Theorem for the first time. The original theorem had already been proved by mathematicians. Anthropic's experiment focused on formalizing that existing mathematical result.
Still, coordinating over 1,000 AI agents to work towards such a demanding technical goal is an interesting demonstration of how AI systems could support complex research.
It Also Designed Three Working Electronic Devices
Anthropic did not test the new models only in mathematics. The company says Fable 5.2 helped design three functional electronic devices in collaboration with electronics manufacturer Diode.
These included:
A handheld virtual pet.
A USB speakerphone.
A home voice-assistant hub.
According to Anthropic, the devices worked on their first attempt.
The experiment highlights an important direction for AI development. Models are increasingly being tested on tasks that go beyond writing text or answering questions. They can assist with product design, hardware planning, software development, and other technical projects that involve multiple connected steps.
This does not mean AI can independently replace engineers or product designers. But it suggests that AI could become a more useful partner throughout the process of turning an idea into a working product.
Mythos 5.2 Is Taking Cybersecurity More Seriously
Cybersecurity is another major focus of this release.
Anthropic describes Mythos 5.2 as its most capable model for cybersecurity engineering so far. It says the model can identify software vulnerabilities, reproduce them, and help researchers develop fixes.
One reported result came from the CVE Rediscovery benchmark, an internal Anthropic evaluation that tests whether a model can rediscover known software vulnerabilities using the affected source code.
Mythos 5.2 scored 89.6% on this evaluation, according to the company.
There is also an interesting comparison with OpenAI's newly launched GPT-6 Cyber.
On Irregular's CyScenarioBench, an external benchmark designed to measure an AI model's ability to plan and execute multi-stage cyber operations, Fable 5.2 completed 76.3% of the challenges. GPT-6 Cyber recorded 61% in the reported comparison.
These figures suggest that Anthropic is taking the cybersecurity competition seriously.
However, benchmark results should be interpreted carefully. They measure performance in specific testing environments and do not automatically translate into success in every real-world security operation.