What happened: Anthropic has released Claude Fable 5.1 and the restricted Mythos 5.1 variant. The update cuts cache-read costs by 75% and adds invisible watermarking on outputs generated after August. The pair continues the dual-track approach the company set with the original Fable 5 / Mythos 5 split earlier this year.
Fable remains the generally available Mythos-class model with safety classifiers in place for cybersecurity, biology, and chemistry prompts. Mythos keeps the same underlying weights with selected safeguards lifted for Project Glasswing partners and other approved users. Pricing for the base pair stays at $10 per million input tokens and $50 per million output tokens.
Key takeaways
- Claude Fable 5.1 and Mythos 5.1 are live as of early September.
- Cache-read costs drop 75% relative to the prior 5.0 generation.
- Post-August outputs carry invisible watermarking.
What the 75% cache cut actually changes
Long-context and agentic workloads lean heavily on cache reads. A three-quarter reduction in that line item is a direct change to the unit economics of running Fable-class models in production. Developers who already moved high-volume jobs onto the 5.0 series now face a lower marginal cost for the same tokens.
The watermarking addition is narrower. It applies to post-August generations and is described as invisible. Anthropic has not published the detection API or the exact robustness claims in the public release notes reviewed so far. For most API users the practical change is the price, not the watermark.
What to watch next
Watch whether the cache pricing shows up immediately on the Anthropic console and third-party routers. Watch for any public detection tools tied to the new watermark. Cross-check against the broader model race: OpenAI’s Astra launch and Google’s Gemini 3.8 Flash both landed in the same window. For the longer creator-stack view see How AI is Reshaping the Content Creator Stack in 2026.