Anthropic’s Fable 5.1 is a bit cheaper, a bit smarter, and refuses a lot less

On Tuesday, Anthropic unveiled the latest iterations of its flagship models, Fable and Mythos. According to the company, these updated versions provide enhanced performance while maintaining lower costs due to reduced cache read pricing. Fable 5.1 and Mythos 5.1 share the same underlying model, but Fable 5.1 is now widely available, while Mythos 5.1 remains available only to Anthropic's trusted users due to specialized cybersecurity and life sciences safeguards.

The price per million input/output tokens remains unchanged at $10/$50, with cache reads now costing $0.25 per million, a 75% reduction. Fable 5.1 is limited to API usage and specific user plans, while Pro subscribers can utilize it through usage credits. Benchmark results show that Fable 5.1 surpasses both the older model and Anthropic's Opus 5, as well as OpenAI's GPT-5.6 Sol.

Anthropic highlights that the model outperforms Fable 5 in most aspects with lower costs, thanks to its ability to decrease reasoning mode by one or two levels without sacrificing performance. One notable improvement is in the Terminal-Bench-Science 0.1 benchmark, where Fable 5.1 scored 52.6% compared to Fable 5's 24.7%. In most other areas, improvements range from 2-4%.

Due to these improvements, Fable 5.1 defaults to high reasoning mode in Claude Code, but medium in Claude Cowork and Claude.ai. Anthropic has also made safeguards more precise, reducing the likelihood of benign content being flagged while still protecting against genuine threats. This change results in the model's safeguards intervening about 60% less per session than with Fable 5, allowing it to identify vulnerabilities in code for coders.

Similarly, the biology safeguards now intervene 85% less often on benign requests about basic biology and medical questions. Anthropic has introduced a watermarking technique for text generated by its models, which identifies the likelihood of Claude writing a given piece of text. While the watermark doesn't affect output quality, it won't contain any user, organization, or conversation data.

A detection API for checking the watermark is now available in private preview for eligible organizations such as regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, and EU civil society groups. Enterprises obligated by the EU's AI Act to verify watermarking will also have access, with a wider rollout planned.

Data retention policy has been a point of contention for enterprises using Fable 5. Now, Anthropic offers Enterprise Frontier Safeguards (EFS) that provide zero data retention while still allowing the company to monitor for misuse. EFS stores data in customer-controlled cloud infrastructure, encrypts it, and runs Anthropic's automated detection.

Any alerts are routed back to the customer. This system will be available to Claude Code, Claude Enterprise, the Claude Platform, and Claude on various cloud platforms, starting this fall. Until EFS is available, eligible customers can use Fable 5.1 with zero data retention.