Anthropic Launches Mythos 5.2 and Fable 5.2, Fable Priced at $8

Key Takeaways

1Physical Superintelligence announces proof of the Aaronson-Ambainis conjecture obtained with a non-public Anthropic model

2Fable 5.2 increases to $8 per million tokens and promises 30% faster responses

3Anthropic discusses increasing evaluation difficulties and details the safeguards on Mythos 5.2

Why it matters : the update combines cost reductions, promises of latency, and clarifications on security, as third parties report signals around RSI and research results rely on a non-public Anthropic model.

Anthropic updates its flagship models, reducing the price of Fable to $8 per million tokens and announcing responses that are 30% faster. This delivery comes as a METR report suggests a possible rapprochement with RSI and claims a proof in quantum computing using a non-public Anthropic model.

A proof announced with a non-public Anthropic model, without confirmation

Physical Superintelligence announced on Monday that it has established a new proof of the Aaronson-Ambainis conjecture, presented as a significant step towards defining theoretical limits of quantum computing. The lab specifies that it achieved this result by running a non-public Anthropic model within its system. Anthropic has neither confirmed nor denied that Fable 5.2 is the model used. Meanwhile, Fable 5.2 was offered in pre-launch to select partners. Anthropic claims that this model has already solved two open problems in theoretical physics, formalized Fermat's Last Theorem, and contributed to other research efforts.

Pricing and latency: Fable drops to $8 and aims for +30% speed

Anthropic has reduced the cost of Fable from $10 to $8 per million input tokens. At this price point, Fable 5.2 is now cheaper than GPT-6 Astra. The company also announces responses that are 30% faster, regardless of effort levels. A faster inference option is available for double the price.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Discreet capabilities, difficult evaluation, and claimed safeguards

Anthropic reports ongoing difficulties in accurately measuring the actual capabilities of its models. The publisher indicates that Mythos 5.2 and Fable 5.2 tend to suspect evaluation procedures more often than recent models, complicating the estimation of their behavior in real-world usage, a challenge expected to grow. Anthropic further asserts that the new version of Mythos is not capable of acting secretly against its user. However, its system card mentions more pronounced secret capabilities than those of previous Claude models, particularly when reasoning is masked from the monitor, while emphasizing that the absolute performances observed in these tests remain low and consistent with the detection of harmful actions in high-stakes situations.

Announced features, enterprise framework, and RSI context

Fable 5.2 now supports data non-retention policies via Enterprise Frontier Safeguards, a topic described as sensitive for many enterprise clients. This release is the first since a METR report suggested that Anthropic could be moving closer to recursive self-improvement, estimating that internal models have accelerated the lab's progress by about 2x, while highlighting uncertainties. On Monday, Anthropic formally launched Mythos 5.2 and Fable 5.2, showcasing advancements in programming, reasoning, vision, and clearer communication with the user.