Cache reads drop 75% with Fable 5.1

Anthropic on September 1 shipped Fable 5.1. For developers the main change is that cached context now costs a quarter of what it did, and the cyber and biology safeguards that kept getting in the way of Fable 5 users now step in much less often.

List prices stay at $10 per million input tokens and $50 per million output tokens. Cache reads are the charge for input the model has already processed and stored, and they drop 75% to $0.25 per million, down from $1. Anthropic’s announcement estimates that typical token-billed workloads come out about 25% cheaper than on Fable 5, and that highly agentic work will often save much more, up to approximately 45%. Mythos 5.1 is the same model with fewer restrictions and stays inside Anthropic’s trusted access program, The New Stack reports, with safeguards Anthropic says are “specifically designed to support work in cybersecurity and the life sciences.” On the biology side, Anthropic has set up an access program with the US government and expects to open enrollment for scientists soon.

Access to Fable 5.1 hasn’t changed. It’s available through the API and on Max, Team Premium and Enterprise Premium plans at 50% of their weekly usage limits. Pro subscribers can reach it with usage credits.

Lower effort, same results

In Anthropic’s charts, Fable 5.1 beats Fable 5, Opus 5 and OpenAI’s GPT-5.6 Sol. The one big jump is on Terminal-Bench-Science 0.1, which tests how well the model drives agents through scientific research: Fable 5.1 scores 52.6% there against Fable 5’s 24.7%. Elsewhere the gains are modest, mostly within 2-4% of Fable 5 or Opus 5.

Anthropic’s pitch is efficiency more than raw score. At Low or Medium effort, it says, Fable 5.1 matches or beats Fable 5 at a much lower cost. So you can drop the reasoning setting one or two notches, get the results you used to get, and use fewer tokens along the way. The defaults reflect this: High effort in Claude Code, Medium in Claude Cowork and on Claude.ai.

What the safeguards still block

At the Fable 5 launch, IBM reported that Anthropic had placed AI-powered classifiers in front of the model, and that these sent prompts touching on cybersecurity, biology, chemistry or model distillation to a less capable model. Anthropic said then that more than 95% of Fable 5 sessions involved no fallback. In practice, the model was still often too cautious and handed requests to Opus whenever it sensed a user pushing its limits.

According to Anthropic, the classifiers are now more precise and “less likely to flag benign content (like queries about medical issues or cyber defenders using the model to make their systems safer).” The cyber safeguards intervene about 60% less per session than with Fable 5, a figure Anthropic also gives as 60% fewer false positives. That’s partly because Fable 5.1 is now allowed to find vulnerabilities in code. Penetration testing, exploit generation and scanning binaries for vulnerabilities are still off limits, and those requests go to the Opus models. The biology safeguards intervene 85% less often on benign questions about basic biology and medicine.

Zero retention, on the customer’s cloud

For regulated businesses, Fable 5 was hard to adopt. Anthropic’s other models offer zero data retention to qualifying customers, but Fable 5 users had to opt into a 30-day retention period. Its replacement is Enterprise Frontier Safeguards (EFS). The data goes to the customer’s own storage on Amazon S3, Azure Blob Storage or Google Cloud Storage, encrypted with the customer’s keys. Anthropic’s automated detection runs against it there, with no human review on Anthropic’s side, and any alerts go back to the customer. Anthropic says the design came out of talks with customers aimed at “the privacy of ZDR and the safety allowed by monitoring across time and accounts.”

EFS covers Claude Code, Claude Enterprise, the Claude Platform, and Claude on Amazon Bedrock, Microsoft Foundry and Google’s Agent Platform. It costs nothing beyond the customer’s own storage bill. The rollout comes in phases starting this fall, and until then Anthropic retains no data from Fable 5 or Fable 5.1 usage.

The release also opens a detection API, in private preview, for Anthropic’s text watermark. The watermark is an invisible numerical marker carried by output from models released after August 2, 2026. Access goes first to organizations eligible under EU law, such as regulators, media, fact-checkers and researchers, and to enterprises that must verify watermarking to comply with the AI Act. A wider rollout is planned.

A cost for agent harnesses

Anthropic is also tightening its defenses against distillation, the practice of training a model on another model’s outputs. New accounts can no longer manually edit Claude’s prior context in a multi-turn conversation while keeping the transcript of its earlier thinking. Anthropic calls this “a common, publicly documented distillation technique.” It’s also what agent harnesses do when they rewrite or compact conversation history, so those tools will break on new accounts. Anthropic says the restriction will reach existing accounts with future model releases.