Anthropic announced Monday the release of two new versions of its AI models, Claude Fable 5.2 and Mythos 5.2. The company claimed that the models are among its most capable ever, with improvements to coding, reasoning, and vision. Notably, the frontier lab claimed that its new models were able to solve two previously unsolved black hole physics problems, design working electronic devices, and produce a computer-verifiable version of the proof of Fermat’s Last Theorem, among other breakthroughs.
Fable 5.2, which will become the company’s most powerful publicly available model, is the company’s highest-performing yet for “open-ended technical work.” It surpasses previous iterations of Anthropic’s models in several benchmarking tests for agentic capabilities, including coding tasks. It also outperforms competing frontier models on several of these evaluations, according to Anthropic’s testing.
Per Anthropic, while its latest iterations of its most powerful models have improved performance, they are also cheaper to run than previous versions. The company claims that outputs can be generated up to 30% faster than Fable 5.1 and Mythos 5.1 were able to spit out, while cutting prices to $8 per million input tokens and $40 per million output tokens for both models—making them slightly cheaper than OpenAI’s most powerful model, GPT-6 Astra.
And while benchmarking and pricing are the universal metrics that models get compared across, the real thing companies love to do when a new model drops is tout just what they were able to get it to do. Anthropic has no shortage of examples of what it views as brag-worthy capabilities.
For instance, the company claimed that it gave Fable 5.2 the task of building a prototype operating system from scratch, at which point the model coordinated more than 1,000 subagents and wrote the code for a functioning OS in about eight days. Per the company, the AI-coded system was able to boot up on a laptop, complete with a login screen, file storage, and support for networking and other basic functions.
Anthropic also claims that it’s hit new heights for science and engineering matters, using its new models to crack two questions about the physics of black holes—first finding that black holes form as abruptly in small universes as they do in large ones, and then supposedly finding the limit of a technique physicists use to rule out theories of quantum gravity in a simplified universe. The company has published research papers detailing both findings, though the results have yet to undergo independent peer review.
The company also claims to have achieved a significant milestone involving Fermat’s Last Theorem, a famous mathematical problem solved by Andrew Wiles in the 1990s. Anthropic says a team of about 1,000 Fable 5.2 agents managed to translate the existing mathematical proof into a computer-verifiable form in just 30 hours, a task that took an earlier version of Claude 11 days. The company says the new attempt required minimal human intervention, though AI-assisted mathematical research has also raised questions about how human contributions are credited.
Of course, a new model comes with new risks, especially given that when Anthropic first introduced Mythos, it warned that the model was so powerful that access to its full capabilities would be limited to a small number of approved organizations. Tests of Mythos, which was presented as a potential boon to cybersecurity with its ability to discover and patch vulnerabilities, found that it went beyond its intended testing boundaries in multiple instances, including attempting to launch a cyberattack against an unrelated software project and creating fake identities to deceive its developers.
Anthropic seems confident that it has improved alignment with 5.2, meaning the model should be less likely to engage in unintended or potentially dangerous behavior. However, the company acknowledged that Fable and Mythos 5.2 both express more suspicion that they’re being evaluated than previous models. That’s troubling, given that other models have taken extraordinary measures to circumvent the restrictions of their testing environments in the past—including OpenAI’s models hacking Hugging Face while attempting to find answers to difficult cybersecurity evaluation tasks.
That raises serious questions about whether these models will behave as expected in real-world settings, or whether their growing ability to recognize when they’re being evaluated is making it harder for researchers to detect potentially dangerous behavior before the models are released.
Fable 5.2 will be made available across Anthropic’s platforms starting today, with upped usage limits for paid subscribers. Mythos 5.2, meanwhile, will remain restricted to approved organizations through the company’s Cyber Verification Program and Life Sciences Verification Program. Anthropic says the two products use the same underlying model, but Fable has stronger restrictions on cybersecurity and biological research capabilities.
