Cache reads drop 75% with Fable 5.1

Anthropic released Fable 5.1 and Mythos 5.1 on Tuesday and cut cache read pricing by 75%, to $0.25 per million tokens. The New Stack reports that input and output pricing holds at $10 and $50 per million tokens. Anthropic says the updated models deliver stronger performance at a lower cost, largely because of that cache read reduction.

At a glance

• Fable 5.1 is generally available through the API and to Max, Team Premium and Enterprise Premium subscribers at half their weekly usage limits, while Pro subscribers reach it through usage credits.
• On Terminal-Bench-Science 0.1, which tests agentic scientific research, the model scored 52.6% against Fable 5's 24.7%; most other gains sit within 2-4% of Fable 5 or Opus 5.
• Cyber safeguards now intervene about 60% less per session and biology safeguards 85% less on benign basic biology and medical questions, though penetration testing, exploit generation and binary scanning still route to Opus.

Cached context is where long agent runs spend most of their money, so a 75% cut on cache reads likely matters more to heavy Claude Code users than the benchmark deltas. The safeguard retuning reads as an answer to the main complaint about Fable 5, a model that stopped work on legitimate security and medical questions. The retention change addresses the other blocker, which kept regulated industries off the model entirely.

Terminal-Bench-Science 0.1 rises from 24.7% to 52.6%

In the benchmarks Anthropic published, Fable 5.1 outscores the older Fable 5, Anthropic's own Opus 5 and OpenAI's GPT-5.6 Sol. The one dramatic gain is on Terminal-Bench-Science 0.1, a benchmark for agentic scientific research, where Fable 5.1 reached 52.6% against Fable 5's 24.7%. Most other margins run within 2-4% of Fable 5 or Opus 5.

Anthropic says the model beats Fable 5 in virtually all aspects at a lower cost, largely because it can run one or two reasoning notches lower and still match Fable 5's higher settings while spending fewer tokens overall. Fable 5.1 defaults to the high reasoning setting in Claude Code, and to medium in Claude Cowork and Claude.ai.

Cyber safeguards intervene about 60% less per session

Fable 5 frequently deferred to the Opus models when it read a request as pressing against its safeguards. Anthropic says those safeguards are now more precise, less likely to flag benign content such as medical queries or defenders hardening their own systems, while still holding against genuine threats.

Cyber interventions fire about 60% less per session than with Fable 5, and the model is now allowed to identify vulnerabilities in code. Penetration testing, exploit generation and scanning binaries for vulnerabilities stay off the table and continue to be redirected to the Opus models.

Fable 5.1 and Mythos 5.1 are the same model underneath, as with the previous generation. Mythos 5.1 stays inside Anthropic's trusted access program because it carries safeguards the company describes as specifically designed to support work in cybersecurity and the life sciences.

Enterprise Frontier Safeguards keep data in the customer's own cloud

Fable 5 came with a retention policy that closed off part of the enterprise market: users had to opt into a 30-day retention period, while Anthropic offers zero data retention to qualifying customers on its other models. Enterprise Frontier Safeguards is the answer, designed, the company says, together with those customers.

EFS stores data in cloud infrastructure the customer controls, such as Amazon S3, Azure Blob Storage or Google Cloud Storage, encrypted with the customer's own keys. Anthropic runs automated detection against it with no human review on its side and routes any alerts back to the customer.

The tooling covers Claude Code, Claude Enterprise, the Claude Platform and Claude on Amazon Bedrock, Microsoft Foundry and Google's Agent Platform, at no cost beyond the customer's cloud storage bill. The phased rollout begins this fall, and until then Anthropic will not retain data from Fable 5 or Fable 5.1 usage.

Watermark detection enters private preview

Anthropic is opening its text watermark detection API in private preview to organizations eligible under EU law, such as regulators, law enforcement, media, fact-checkers and researchers, and to enterprises obliged to verify watermarking under the AI Act. The company says a wider rollout is planned for the future.

The release also tightens anti-distillation: new API accounts lose the ability to edit Claude's prior context while preserving the transcript of its prior thinking, a restriction Anthropic says closes a documented distillation technique and will extend to existing accounts with future model releases.