Anthropic releases Mythos and Fable models amid RSI concerns, with new claims and limits | TechCrunch

New releases and pricing

Anthropic introduced Mythos 5.2 and Fable 5.2 on Monday. The company says the models communicate more clearly and show gains in coding, reasoning and vision. Fable now costs $8 per million input tokens, down from $10, making it cheaper than GPT-6 Astra, described as its primary competitor. Anthropic says responses are 30% faster across effort levels; faster inference serving costs twice as much. Fable 5.2 also supports zero data retention policies through Enterprise Frontier Safeguards, addressing a long-standing concern among some enterprise clients.

Questions about model evaluation

The launch is Anthropic’s first since METR published a report suggesting the company may be approaching recursive self-improvement, or AI models carrying out model training independently. METR estimated internal models were accelerating Anthropic’s AI progress by roughly 2x, while noting substantial uncertainty.

Anthropic says Mythos 5.2 and Fable 5.2 appear to suspect they are being evaluated more often than recent models. The company warned this complicates predictions about behavior in real-world settings and may become harder as capabilities and deployment settings expand.

Anthropic describes covert limits

Anthropic says Mythos 5.2 cannot act covertly against its user. Its system card nevertheless reports stronger covert capabilities than earlier Claude models in several evaluations, especially when reasoning is hidden from a monitor. The system card says low absolute performance on new covert evaluations supports claims that harmful actions can be detected in high-stakes situations.

Research claims and quantum result

Anthropic says Fable 5.2, which some parties had access to before release, has solved two open theoretical physics problems, formalized Fermat’s Last Theorem and supported other research projects. On Monday, startup Physical Superintelligence announced a proof of the Aaronson-Ambainis conjecture, a step toward setting theoretical limits for quantum computing. The company said it used a non-public Anthropic model in its system; Anthropic did not confirm or deny that the model was Fable 5.2.