Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, pairing lower token prices with different access conditions for security work. Both use the same underlying model.
Standard pricing is $8 per million input tokens and $40 per million output tokens. Optional fast mode doubles those prices to $16 and $80. Fable is broadly available; Mythos access is verified.
Anthropic reports improvements in coding, reasoning and vision. Those are company findings, rather than an independent guarantee of better results. It also acknowledges that models increasingly recognising evaluations makes real-world behaviour harder to assess.
Security access depends on the work
The Cyber Verification Program documentation describes three tiers. Individuals can apply for Defense Access, covering work such as incident response and vulnerability triage. Red Team Access, which covers authorised penetration testing and exploit development, is for organisations. Approval for internal work does not automatically authorise a customer-facing service: that requires separate approval.
That distinction matters when choosing a model for a security workflow. A model name and an approved account do not establish that every task or client engagement falls within the granted access.
Privacy still includes monitoring
Anthropic’s Enterprise Frontier Safeguards architecture, first announced in September, stores monitoring data in infrastructure controlled by the customer. Automated review can examine activity across requests, with flags routed to the customer’s own reviewers. The design aims to support zero data retention by Anthropic while preserving misuse detection. The November launch announcement says EFS remains in private beta.
For an enterprise, the practical questions remain who stores activity, who controls encryption keys, and who reviews an alert. A zero-retention arrangement with a model provider does not mean that an organisation keeps no logs.
Acceleration is not the same as autonomous self-improvement
The release follows a METR assessment finding likely acceleration in Anthropic’s development, primarily through AI assistance. The assessment carries substantial uncertainty, and important evidence was redacted. As DaySec previously reported, that evidence needs careful interpretation. It does not establish that these newly released models independently run an entire model-development process.
