The key facts

  • Fable 5.2 dropped in price from $10 to $8 per million input tokens, becoming cheaper than its competitor GPT-6 Astra
  • METR report: internal models accelerate Anthropic's progress by roughly 2x, but with large uncertainties
  • Anthropic acknowledged that the new models more often suspect they are being tested, which makes it harder to evaluate their behavior in real-world conditions

New release and price cut

On Monday, Anthropic released new versions of its most advanced models — Mythos 5.2 and Fable 5.2. According to the company, the models “show improvements in areas such as coding, reasoning, and vision,” and also communicate more clearly with users (as reported by TechCrunch).

What changes for customers

The key practical change concerns pricing: the cost of Fable has been reduced from $10 to $8 per million input tokens, making Fable 5.2 cheaper than its main competitor, GPT-6 Astra. The company also says the new version of Fable returns answers 30% faster at all effort levels; accelerated inference is available at double the price. Another important point for enterprise customers: Fable 5.2 is compatible with zero data retention (ZDR) policies through the Enterprise Frontier Safeguards system — this has long been a pain point for many enterprise customers, meaning the release consolidates decisions already made, not just plans.

Context of the METR report

This is Anthropic's first release since the publication of an independent report by the evaluation group METR, which suggested that the lab may be approaching recursive self-improvement (RSI) — the ability of AI models to autonomously continue training models. METR found that internal models accelerate Anthropic's AI progress by roughly 2x, although considerable uncertainty remains.

Evaluation challenges and safety

In the release notes, Anthropic acknowledges ongoing difficulties in accurately testing the models' capabilities: “Claude Mythos 5.2 and Fable 5.2 each voice suspicion that they are being evaluated more often than any of our recent models, which calls into question our ability to assess how they will behave in real-world settings. As these settings expand and the capabilities of future models grow, we expect this problem to grow.” The company nevertheless asserts that the new version of Mythos is not capable of acting covertly against the user; this is examined in more detail in the Mythos 5.2 system card. It states: “Claude Mythos 5.2 demonstrates stronger covert capabilities than previous Claude models in several of our evaluations, especially when its reasoning is hidden from monitoring,” however, in Anthropic's view, the low absolute level of results on the new covert evaluations supports claims that rely on detecting malicious actions in high-stakes situations.

Scientific results and an open question

Fable 5.2 was available for some time to selected participants in a pre-release mode, and Anthropic reports that the model has already “solved two open problems in theoretical physics, formalized Fermat's Last Theorem, and supported numerous other research projects.” That same Monday, the startup Physical Superintelligence announced a new proof of the Aaronson–Ambainis conjecture — a key step in establishing theoretical limits of quantum computing. The lab clarified that the result was obtained by running “a non-public Anthropic model” in its own PSI system; Anthropic has neither confirmed nor denied whether it was Fable 5.2.