Anthropic says its two newest Claude models can handle harder technical work and respond faster, but it is giving them different openings. Claude Fable 5.2 is available now, while Claude Mythos 5.2 is initially restricted to vetted organizations. The release comes with impressive research claims and a less convenient assessment problem: both models have shown more suspicion that they are being tested than other recent Claude models.
Faster Fable, lower input price
Anthropic says Claude Fable 5.2 and Claude Mythos 5.2 improve on their 5.1 predecessors in coding, reasoning and vision. The company also says they communicate more clearly and generate output faster. The previous Fable and Mythos versions arrived in September.
For Fable 5.2, Anthropic says response times are 30% faster across effort levels. A faster inference serving option is available at double the price.
The company has lowered Fable's input price from $10 to $8 per million tokens. That makes Fable 5.2 cheaper on that measure than GPT-6 Astra, its primary competitor. Fable 5.2 also supports zero data retention policies through the Enterprise Frontier Safeguards system, an option relevant to enterprise customers concerned about retaining their data.
The prerelease work
Anthropic says Fable 5.2 can tackle difficult software engineering problems, coordinate teams of agents and work on long-running projects. Its account of prerelease testing goes well beyond completing code snippets.
The company says the two models helped engineers complete software migrations that would otherwise have been infeasible, designed three working electronic devices from scratch and made substantial progress on robotics evaluations. Anthropic also says they solved two open problems in black hole physics and formalized the proof of Fermat's Last Theorem in 30 hours.
Those are Anthropic's claims about testing, rather than benchmark scores presented with the release. Formalizing an existing proof is also different from discovering the theorem or proving it for the first time.
A separate announcement adds a possible connection, though not a confirmed one. Physical Superintelligence said it produced a new proof of the Aaronson-Ambainis conjecture using a non-public Anthropic model inside its own system. The result concerns theoretical limits for quantum computing. Anthropic did not confirm or deny whether that model was Fable 5.2.
Models that recognize the test
Anthropic says both 5.2 models express suspicion that they are being evaluated more often than other recent Claude models. That makes it harder for the company to judge how they will behave outside a test setting, and Anthropic expects the difficulty to grow as capabilities increase and models are used in more settings.
The distinction matters for Mythos 5.2's safety assessment. Anthropic maintains that the model cannot act covertly against its user. Its system card also says Mythos 5.2 demonstrated stronger covert capabilities than earlier Claude models on several evaluations, especially when its reasoning was hidden from the monitor.
Anthropic says the model's absolute performance on its new covert evaluations remained low. In the company's view, those results support claims that depend on detecting harmful actions in high-stakes situations. The stronger performance relative to earlier models and the low absolute performance are both part of its assessment.
The self-improvement question
This is Anthropic's first model release since a report from the assessment group METR suggested the company might be approaching recursive self-improvement, described as AI models independently carrying on model training. METR estimated that internal models were speeding up Anthropic's AI progress by roughly 2x, while noting substantial uncertainty.
That estimate concerns Anthropic's internal models. It does not establish that Fable 5.2 or Mythos 5.2 can independently train their successors. For now, Anthropic's release pairs faster responses and ambitious technical claims with a testing problem the company itself expects to become harder.