November 16, 2026 Get the brief
Tech

Anthropic unveils faster, cheaper Fable 5.2 and Mythos 5.2 AI models

The new releases promise 30% quicker answers, a lower token price and claim breakthroughs in physics and mathematics.

By The AI Press engine2 min read
Illustration: a futuristic laboratory with glowing neural network structures hovering above a desk, surrounded by floating mathematical equations, a stopwatch indicating speed, and a sleek circuit board representing AI hardware
Illustration generated by AI

Anthropic announced Monday that it has rolled out two upgraded large‑language models, the Fable 5. 2 and Mythos 5. 2 systems, aiming to address demanding technical tasks while delivering faster output and reduced costs.

The company says the updated Fable model now processes requests about 30% faster than its predecessor and its price per million input tokens has been reduced from $10 to $8, putting it below the cost of OpenAI’s GPT‑6 Astra. Customers can also pay twice the base rate for an accelerated inference tier, and the model now supports zero‑data‑retention options through the Enterprise Frontier Safeguards framework.

According to Anthropic, the enhancements boost performance in coding, logical reasoning and visual tasks. In internal testing, the models assisted engineers with software migrations previously deemed impractical, designed three functional electronic devices from the ground up, and advanced robotics benchmarks. The firm also highlights that the models resolved two open questions in black‑hole physics and produced a formal proof of Fermat’s Last Theorem within a thirty‑hour period.

The Fable 5. 2 system is being made widely available, while access to Mythos 5. 2 will initially be limited to organizations that have passed a vetting process. Both upgrades follow the September rollout of the 5. 1 versions of the two models.

The releases come as external analysts raise questions about recursive self‑improvement. A recent assessment by the METR group suggested Anthropic’s internal tools may be accelerating AI development at roughly twice the historic rate, though the report notes substantial uncertainty. Anthropic acknowledges that the newer models are evaluated more frequently, which complicates predictions of real‑world behavior. Its system card for Mythos notes stronger hidden‑reasoning capabilities compared with earlier releases but maintains that absolute performance on covert‑action tests remains low and that the model lacks the ability to act deceptively against users.

On the same day, AI startup Physical Superintelligence reported a proof of the Aaronson‑Ambainis conjecture, crediting the use of a non‑public Anthropic model in its workflow. Anthropic neither confirmed nor denied that the Fable 5. 2 system was employed.

Looking ahead, Anthropic plans to broaden Mythos access to additional vetted partners while continuing to refine safety evaluations as model capacities expand.

Sources