Anthropic announced on Monday (9) two new versions of its artificial intelligence (AI) models: Claude Fable 5.2 and Claude Mythos 5.2. The tools arrive with advances in programming, reasoning, and visual interpretation, but the launch comes amid an important debate about AI's ability to accelerate its own development.
According to the company, the new models can communicate more clearly and provide faster responses. Fable 5.2 also had a price reduction: the entry cost fell from US$ 10 to US$ 8 per million tokens (about R$ 43), a 20% decrease. Anthropic says that responses became 30% faster at different levels of effort.
Mythos 5.2 will have restricted access to organizations verified by the company, including participants in programs focused on cybersecurity.Meanwhile, Fable 5.2 is available through Anthropic’s API and cloud computing platforms.
AI may be accelerating its own development
The launch comes just a few days after an independent study by the AI security organization METR found that Anthropic’s models are evolving almost twice as fast as before, partly due to the use of AI in the development process itself.
The analysis, conducted over 11 weeks and published on November 5, estimated that the company’s model performance on well-defined tasks had advanced 1.9 times faster since October 2025 than during the preceding period.
The result sparked renewed discussions about so-called recursive self-improvement (RSI, for its English acronym). The concept describes the possibility of AI systems helping to develop more advanced versions of themselves, accelerating successive generations of technology.
There are, however, important caveats. METR itself characterized the estimate as highly uncertain: the actual acceleration could be between 1.5 and 2.5 times the historical pace. The study also emphasized that AI still cannot make highly significant contributions in a fully autonomous manner during the most important stages of model training.
According to the company, the new models can communicate more clearly and provide faster responses – Image: jackpress/ShutterstockRead more:
- Investors doubt Anthropic will be worth R$8.19 trillion in its IPO
- Fall in tech stocks threatens Anthropic’s stock market debut
- Anthropic invests billions in gas generators for data centers
More capable models are also harder to evaluate
- Anthropic acknowledged that the new models pose challenges to evaluation processes;
- According to the company, both Mythos 5.2 and Fable 5.2 suspected more frequently than previous versions that they were being tested;
- This behavior may make it harder to predict how the tools will act in real-world situations;
- The company expects the issue to become more significant as the models gain capabilities and are used in more complex environments;
- The Mythos 5.2 safety card also points to advances in covert action capabilities in certain evaluations, especially when the model's reasoning is hidden from monitoring systems;
- Anthropic, however, says absolute performance on these tasks is still low and maintains that the model cannot act covertly against the user.
AI has already helped solve scientific problems
Anthropic says Fable 5.2 participated in scientific research even before its public launch.Among the results attributed to the model are the resolution of two open problems in theoretical physics and the formalization of a proof of Fermat’s Last Theorem.
The company also reports that its models helped engineers complete software migrations considered unfeasible, design three functional electronic devices from scratch, and advance evaluations related to robotics.
In another announcement made this Monday, the startup Physical Superintelligence released a new demonstration related to the Aaronson-Ambainis conjecture, an important problem for establishing the theoretical limits of quantum computing. The company said it obtained the result using a non-public Anthropic model, but the developer did not confirm whether the tool was Fable 5.2.
Rodrigo Mozelli is a journalist graduated from São Paulo Methodist University (UMESP) and is currently a reporter at Olhar Digital.
