Anthropic shipped Claude Fable 5.1 on 1 September. Base token prices did not move: still $10 per million input and $50 per million output. What changed is cache reads, cut 75% to $0.25 per million tokens. Anthropic says that makes typical token-billed workloads about 25% cheaper than Fable 5, and highly agentic runs (long tool loops, ads desks that keep context warm) up to around 45%.

Kill the headline percentage. “Up to 45%” is Anthropic’s cache-heavy agentic scenario, not a guarantee on your bill. Artificial Analysis ran the model and found Fable 5.1 can still cost more per task, about 20% more than Fable 5 at max effort on their Intelligence Index, because it burns roughly 1.7× the output tokens. The cache cut saves real money on those loops; output volume can eat it. Reprice your stack with your own traces, not the keynote slide.

For a Spanish agency or EU SMB already wiring Claude into Meta ads tooling, CRM triage or long Cursor sessions: map where cache hits actually land. If your agent re-reads the same account context every turn, the $0.25 cache line matters. If every run dumps fresh long output, you may not feel the “25% cheaper” story.

Enterprise Frontier Safeguards are the retention piece. Anthropic says EFS stores data in customer-controlled cloud infrastructure (ZDR-like privacy with misuse detection still in place), rolling later this fall on Claude Code, Enterprise, the API and the big cloud partners. Until then, eligible customers can run Fable 5.1 with zero data retention. That is the lever for Spanish clients who will not park CRM dumps on US retention defaults. Ask for the eligibility path before you put production customer data through a new model.

The EU bit is Article 50. Fable models released after 2 August ship an invisible watermark under the EU AI Act Code of Practice on transparency. Anthropic says it does not change output quality and carries no user or conversation metadata. A detection API is in private preview for obligated organisations (regulators, media, enterprises that must verify watermarking). If you publish client copy or ads that Claude drafted, ask now who in the account needs detection access, and whether your contracts already require disclosure of AI-assisted content.

Mythos 5.1 is the same model with looser cyber/biology safeguards. Access is US-gated first through trusted programs. Do not plan a Spanish life-sciences or red-team workflow on Mythos until your region is in.

This week’s job is narrow: pull last month’s Claude token mix (cache vs output); model Fable 5.1 with both Anthropic’s cache story and higher output; confirm ZDR/EFS eligibility for any client CRM path; write one sentence into your AI-content SOP about Art. 50 watermark and who holds the detection key. Cheaper agent loops are useful. Unaudited bill savings and unmarked client copy are not.