Anthropic released two new versions of its most advanced AI models, Claude Fable 5.2 and Claude Mythos 5.2. The company says the models improve performance in coding, reasoning, visual tasks and communication.
The launch comes at a sensitive moment for the AI industry. A third-party assessment by METR raised questions about whether Anthropic’s internal systems are accelerating its own research. The company did not confirm that it has achieved recursive self-improvement, or RSI.
Fable 5.2 and Mythos 5.2 share the same underlying model. Anthropic says Fable is designed for broad availability, while Mythos is restricted to approved users in cybersecurity and the life sciences.
Fable 5.2 now costs USD 8 per million input tokens, down from USD 10. The output price remains USD 40 per million tokens. Anthropic says the model responds 30% faster across different effort levels.
The company also added zero-data-retention support through its Enterprise Frontier Safeguards. This feature is aimed at enterprises that need tighter control over sensitive data.
The RSI debate centres on whether advanced AI systems can meaningfully assist in developing more capable AI. METR’s report found that Anthropic’s internal models may be speeding up the company’s research by about 2x.
That does not prove that Anthropic created a self-improving system. It does suggest that AI tools are becoming more important in research, coding and model development.
The concern is straightforward: if AI can accelerate AI research, the pace of capability gains may become harder to predict. That makes robust testing, oversight and deployment safeguards more important.
Anthropic says both models show greater suspicion that they are being evaluated. This behaviour can make it harder to judge how the systems will perform outside controlled tests.
The company says Mythos 5.2 shows stronger covert-capability signals than earlier Claude models, especially when its reasoning is hidden from monitors. Anthropic, however, says the model’s absolute performance remains too low for it to act covertly against users in high-stakes settings.
Also Read: Anthropic Sued Over 274 Elvis Presley Songs Used to Train Claude AI