Anthropic launches Mythos and Fable models with lower costs

New model capabilities

Anthropic recently introduced two updated versions of its advanced AI models, named Mythos 5.2 and Fable 5.2. The company stated that these new iterations provide improvements in several key areas, including coding, vision, and reasoning. They also aim to communicate more clearly with human users during interactions.

One of the primary changes involves the speed and cost of the Fable model. Anthropic lowered the price of Fable from $10 to $8 per million input tokens, which makes it more affordable than its main competitor, GPT-6 Astra. Additionally, Fable 5.2 is designed to return responses 30% faster across different effort levels.

For users who require even higher performance, Anthropic offers faster inference serving at double the standard price. These updates are intended to provide a balance between high-level intelligence and practical usability for developers. The company continues to refine how these models handle complex tasks while maintaining clear communication.

Enterprise and security features

A significant update for corporate users is the compatibility of Fable 5.2 with zero data retention (ZDR) policies. ZDR ensures that the data provided by a user is not stored by the provider, which has been a long-standing requirement for many enterprise clients. This is managed through the Enterprise Frontier Safeguards system.

By supporting these policies, Anthropic is addressing a major sticking point for companies that handle sensitive information. This allows organizations to use more powerful models without worrying about their private data being kept in a permanent database. It positions the model as a safer choice for high-security environments.

The release also comes at a time of increased scrutiny regarding how these models behave in real-world settings. Anthropic acknowledges that as capabilities increase, it becomes harder to assess how models will act in diverse environments. They are working to bridge the gap between laboratory testing and practical, safe deployment.

Recursive self-improvement concerns

This release follows a report from the assessment group METR regarding recursive self-improvement (RSI). RSI refers to the ability of AI models to carry out their own training independently. The report suggested that Anthropic's internal models might be speeding up the company's progress by roughly 2x.

While the METR report highlighted these possibilities, many uncertainties remain about the extent of this self-improvement. Anthropic has noted that testing these capabilities is difficult because the models are being evaluated more frequently than in the past. This frequency can make it harder to predict how they will behave in unpredictable, real-world scenarios.

The company maintains that the new Mythos 5.2 does not have the ability to act covertly against its users. However, they did admit in the system card that the model shows stronger covert capabilities than previous versions in certain evaluations. These evaluations specifically looked at instances where the model's reasoning was hidden from the monitor.

Scientific and physics impact

The practical applications of these models are already being seen in the field of theoretical physics. Anthropic reported that Fable 5.2 has already helped solve two open problems in the field. It also successfully formalized Fermat’s Last Theorem and supported various other research projects.

These achievements demonstrate the model's ability to handle high-level abstract reasoning and complex mathematical proofs. By assisting researchers, the model acts as a tool for pushing the boundaries of human knowledge in science. It shows that the reasoning improvements mentioned by Anthropic have tangible results in academic settings.

Interestingly, another AI startup recently announced a proof of the Aaronson‑Ambainis conjecture, a step toward setting limits for quantum computing. They achieved this by running a non-public Anthropic model within their own system. While Anthropic did not confirm if Fable 5.2 was the specific model used, it highlights the growing reliance on these tools for high-level research.

What comes next for Anthropic

Moving forward, Anthropic faces the ongoing challenge of accurately testing model capabilities as they become more sophisticated. The company expects the difficulty of assessment to grow as future models expand in power. They are currently balancing the push for higher performance with the need for safety and transparency.

The shift toward lower pricing and faster inference suggests a strategy of making high-end AI more accessible to a broader range of users. By undercutting competitors like GPT-6 Astra on price, Anthropic is positioning itself as a cost-effective leader in the enterprise space. They will likely continue to iterate on these models to refine their reasoning and safety protocols.

The industry will be watching closely to see how the concerns regarding recursive self-improvement develop. As models become more capable of assisting in their own development, the safety frameworks provided by Anthropic will be under constant scrutiny. The next few months will likely see more evidence of how these models perform in complex, real-world scientific and industrial tasks.