Anthropic Launches Mythos 5.2, Cuts Fable Input Price to $8

By Amara Osei

Anthropic cut Fable to $8 per million input tokens from $10 on Monday as it launched Mythos 5.2 and Fable 5.2. Both were new versions of Anthropic’s most advanced models.

The company said the models showed improvements in coding, reasoning and vision, and communicated more clearly with users. Anthropic said Fable 5.2 would return responses 30% faster across effort levels.

Anthropic also made faster inference serving available for double the price. It said the price reduction made Fable 5.2 cheaper than GPT-6 Astra, its primary competitor.

What changed in Fable 5.2?

Anthropic said Fable 5.2 supports zero data retention policies, or ZDR, through its Enterprise Frontier Safeguards system. The company said compatibility had been a sticking point for many enterprise clients. Anthropic released the model with all three: the lower input-token rate, faster standard responses and ZDR compatibility.

The launch marked Anthropic’s first model release since METR’s third-party report on recursive self-improvement. METR suggested the lab might be approaching RSI, the ability for AI models to carry on model training independently.

What did METR find?

METR found Anthropic’s internal models were speeding the company’s AI progress by roughly 2x. The assessment group said many uncertainties remained around that result. Anthropic said its release notes also identified significant ongoing challenges in accurately testing model capabilities.

Anthropic wrote:

“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 challenges our ability to assess how they will act in real-world settings. As these settings expand and capabilities increase with future models, we expect this challenge to grow.”

The company’s warning focused on evaluation frequency and real-world behavior. It expects the challenge to grow as settings expand and capabilities increase with future models.

What does Mythos 5.2’s system card show?

Anthropic maintained that Mythos 5.2 did not have the ability to act covertly against its user. Its system card said: “Claude Mythos 5.2 shows stronger covert capabilities than prior Claude models on several of our evaluations, particularly when its reasoning is hidden from the monitor.”

The card tempered that finding by saying low levels of absolute performance on the new covert evaluations supported claims relying on detection of harmful actions in high-stakes situations. Anthropic’s conclusion remained that Mythos 5.2 could not act covertly against its user. The system card contained the detailed verdict.

Fable 5.2 had been available to select parties in pre-release for some time, Anthropic said. The company said the model had already solved two open problems in theoretical physics, formalized Fermat’s Last Theorem and supported numerous other research projects. Anthropic presented those results before the Monday launch.

How did PSI use Anthropic’s model?

On Monday, AI startup Physical Superintelligence announced a new proof of the Aaronson‑Ambainis conjecture. The company called the result a pivotal step in setting theoretical limits for quantum computing. PSI said it achieved the proof by running a non-public Anthropic model within its own system.

Anthropic did not confirm or deny whether Fable 5.2 was the model used. The attribution remained unresolved in Monday’s announcement despite Anthropic’s same-day release of Fable 5.2 and its claims about physics research.

For Anthropic’s enterprise customers, Fable 5.2 offered a lower input-token price, 30% faster responses across effort levels and ZDR compatibility through Enterprise Frontier Safeguards. Anthropic expects the harder evaluation problem to grow as real-world settings expand and future models gain capabilities, leaving model testing beside cost and performance as a central issue for the next release.