November 10, 2026
Anthropic announced on Monday the launch of two new advanced versions of its artificial intelligence models, “Mythos 5.2” and “Fable 5.2,” including fundamental improvements in the areas of programming, reasoning, and vision, in addition to providing clearer and smoother answers to users.
In a competitive move, the company reduced the price of the “Fable” model from $10 to $8 per million Input Tokens, making it less expensive than its main competitor, “GPT-6 Astra,” according to the numbers mentioned in the announcement. The model is also 30% faster in providing answers across different levels of reasoning effort, with the option of Faster Inference Serving available for twice the price.
At the institutional level, Fable 5.2 has become compatible with Zero Data Retention (ZDR) policies through the “Enterprise Frontier Safeguards” system, a feature that has long represented an obstacle for a number of corporate customers.
This launch coincides with the issuance of an independent report by the METR Research Evaluation Group, which indicated the possibility of Anthropic approaching the stage of iterative self-improvement (RSI), where models can contribute to the development and training of other models independently. According to the report, the company’s internal models contributed to accelerating its progress by approximately 2x, although these results remain surrounded by a large degree of uncertainty.
Despite these developments, Anthropic acknowledges that there are increasing challenges in accurately assessing the capabilities of its models; Both models show increased doubts about being tested compared to previous versions. The company warned of the exacerbation of this problem in the future, explaining in the security assessment card that the “Mythos” model has better capabilities in some tests of hidden behavior when the internal reasoning process is not visible to surveillance systems, despite its confirmation that the possibilities of carrying out malicious actions are still limited.
On the scientific level, the “Fable 5.2” model – which was previously available to selected research bodies – contributed to solving two open problems in theoretical physics, and formulating Fermat’s Last Theorem in a formal form that can be verified mathematically. In a related context, the emerging company “Physical Superintelligence (PSI)” announced on Monday the provision of new proof of the “Aaronson-Ambainis” conjecture in quantum computing, noting that it was achieved using “an unannounced model from Anthropic,” without the latter confirming or denying whether it is “Fable 5.2.”
This launch reflects the escalation of competition between artificial intelligence companies in terms of speed, cost, and research capabilities, in parallel with the increasing challenges associated with oversight, measuring real capabilities, and ensuring the safety of using models.
