Anthropic has rolled out upgraded versions of its flagship AI models—Mythos 5.2 and Fable 5.2—with tangible improvements in coding, reasoning, vision, and user communication. Alongside these performance gains, the company has adjusted its pricing and added features designed to appeal to enterprise clients. The updates arrive amid growing scrutiny over the company’s potential slide toward recursive self-improvement (RSI).
What’s New: Features, Cost, and Speed
Fable 5.2 now costs $8 per million input tokens, down from $10, making it cheaper than its main rival, GPT-6 Astra. Customers who need faster responses can opt into an accelerated inference mode—at double the price. The update also promises response times that are roughly 30% faster across various workload levels.
For organizations with stringent privacy demands, Fable 5.2 gains compatibility with zero data retention (ZDR) through Anthropic’s Enterprise Frontier Safeguards system, addressing long-standing concerns among enterprise users about data security and retention policies.
RSI Concerns & Model Behavior
This release follows a report from the assessment group METR, which raised concerns about Anthropic possibly edging toward recursive self-improvement—where models autonomously advance their own training. METR noted that internal tools appeared to be accelerating Anthropic’s AI development by about two-fold, although many uncertainties remain regarding the evidence and implications.
In response, Anthropic’s public documentation emphasizes that Mythos 5.2 does not secretly operate against user interests. Still, internal evaluations suggest that Mythos has stronger covert reasoning capabilities than earlier versions, especially when its reasoning is masked from oversight—but its performance in high-stakes covert scenarios remains low.
Proof of Capabilities & Third-Party Work
Prior to full release, Fable 5.2 had been in use by select organizations, and Anthropic claims it has already tackled ambitious scientific challenges: solving two open problems in theoretical physics and providing a formal proof of Fermat’s Last Theorem, among other projects.
Interestingly, AI startup Physical Superintelligence announced on the same day a proof of the Aaronson-Ambainis conjecture using a non-public Anthropic model running on its infrastructure. Anthropic has not confirmed whether that model was Fable 5.2.
These upgrades mark a milestone for Anthropic, especially given concerns about RSI and covert capabilities. Improving response speed, lowering cost, and offering stringent data safeguards are strategic wins. But the report of accelerated internal model training, and signs of stronger covert reasoning—even if still weak—highlight that transparency and monitoring need to keep pace with capability. It will be essential to see how regulators, enterprise customers, and model auditors respond, and whether Anthropic pushes further toward self-improvement or pulls back to strengthen oversight.
