Claude Fable 5.1 and Enterprise Frontier Safeguards: the retention wall comes down, and the bill with it

What changed

On 1 September 2026 Anthropic released Claude Fable 5.1 and Claude Mythos 5.1. They are the same model with different safeguards: Fable 5.1 is generally available on the Claude API as claude-fable-5-1, on Amazon Web Services, Google Cloud and Microsoft Azure; Mythos 5.1 is limited to vetted US organisations through Anthropic's Cyber Verification and Life Sciences Verification programmes. Fable 5.1's list price is unchanged at $10 per million input tokens and $50 per million output tokens, but cache reads, where the model reuses context it has already processed, fall by 75% to $0.25 per million tokens wherever usage is billed by token. Anthropic estimates this cuts the cost of typical Fable workloads by around 25% and highly agentic workloads by up to around 45%, measured on four weeks of its own August usage.

The same day Anthropic announced Enterprise Frontier Safeguards (EFS). In June, Fable 5 and Mythos 5 arrived as Covered Models with a mandatory 30-day retention of prompts and outputs on every surface, including for organisations that otherwise held zero data retention agreements. Under EFS the monitoring still happens, but the retained data is stored in the customer's own cloud account, for example Amazon S3, Azure Blob Storage or Google Cloud Storage, under customer-managed encryption keys and the customer's own access policies. Automated systems review a rolling window for serious misuse; flags go to the customer, and no Anthropic human review is required. Each control is opt-in, none changes model behaviour, pricing or rate limits, and Anthropic does not charge for it. EFS rolls out in phases from autumn 2026 across Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Google's Agent Platform and Microsoft Foundry. Until it arrives, eligible customers get the option of zero data retention on Fable 5 and 5.1 for their own internal business applications, for a limited time, by notice from Anthropic or their cloud provider or by request form. Consumer plans and non-eligible organisations stay under the June policy. The Help Center lists Fable 5.1 and Mythos 5.1 as Covered Models designated on 31 August 2026.

Why it matters for UK business

The retention change is the one to act on. In June, a number of UK firms in regulated or client-confidential work concluded that a Mythos-class model could not sit inside their data-protection posture, because their DPIA could not accommodate a processor holding 30 days of prompts that might contain client material. Anthropic's own account is candid: it says enterprises understood the security case but many, especially in regulated industries, found the policy difficult to use, and that it spent hundreds of hours with more than 100 customers designing the alternative. Reuters reported the direction of travel on 20 August; OpenAI had previewed its own zero-retention approach a day earlier. This was a market correction, and Anthropic has now made it.

What a UK firm gets is not "no retention". It is a choice of three positions, and the DPIA should name which one applies. Position one is the default: 30-day retention held by Anthropic, automated review, controlled-access human review on flags, tamper-evident access logs. Position two is interim zero data retention on Fable 5 and 5.1 for internal applications, available only to eligible organisations, only for a limited time, and only if Anthropic or the cloud provider has confirmed it. Position three is EFS when it reaches you: retention continues, but in your account, under your keys, with your people clearing the flags. For UK GDPR purposes the third position is the interesting one, because the storage limitation and security questions a DPIA asks about a processor become questions about your own cloud estate, which you already answer for everything else. Anthropic has not published region detail for EFS storage; what it has said is that the account is yours, and your account's region policy follows from that.

The cost change is real but conditional. The headline did not move, so a model-selection table that lists Fable at $10 and $50 remains correct. What changed is the price of reading cached context, and the benefit lands in proportion to how much of your traffic is cached: long system prompts, large documents held across turns, and tool-heavy agent loops benefit most; short one-off requests barely notice. Anthropic's 25% and 45% are estimates on its own usage mix. Yours will differ, and the honest way to quote a saving is to re-run your own model with the new rate.

Three smaller items deserve a line each. Fable 5.1 may now be used to identify software vulnerabilities, though not to develop exploits, and Anthropic reports around 60% fewer safeguard interventions per Claude Code session, which matters for development teams who found Fable 5 interrupted defensive work. New API accounts created from 1 September can no longer edit Claude's prior context in a multi-turn conversation while preserving its prior thinking transcript, an anti-distillation measure that Anthropic says will affect a small number of custom integrations and will extend to all accounts with future releases. And because Fable 5.1 was released after 2 August 2026, its text output carries the EU AI Act transparency watermark that Anthropic committed to under the Code of Practice; it carries no user or organisation information, and a detection API is in private preview for eligible bodies.

What to do, and what not to do

Do:

• Submit the EFS request form now if 30-day retention is what kept Fable out of your estate in June. Eligibility is by notice, and the interim zero-retention option is explicitly time-limited.
• Update the DPIA and the record of processing to name the position you are actually in: default retention, interim zero retention, or EFS. They are three different descriptions of the same processor.
• Re-run your cost model with cache reads at $0.25 per million tokens and your own cached-token share. Quote the result, not the vendor's percentage.
• If you run custom multi-turn integrations on a new API account, test them against the prior-context change before moving production traffic to 5.1.

Do not:

• Tell a client or a regulator that Anthropic "no longer retains data". The default is unchanged; only the eligible and the EFS-enabled are different.
• Assume Mythos 5.1 is coming to your UK environment. It is US-only for now, and Anthropic says it is coordinating with the US government on wider access.
• Treat the 25% as a budget line. It is the vendor's estimate on the vendor's traffic.

Where The AI Consultancy fits

Choosing the retention position and evidencing it in a DPIA is the governance half of a Claude deployment, and it is what our Claude implementation service settles before anything is wired up. Our article on whether Claude is GDPR-compliant for a UK business carries a dated note on this change, the residency mechanics are in Claude data residency for UK firms, and the sterling arithmetic is in what Claude costs a UK business.

Verified on 5 September 2026 against Anthropic's announcements "Introducing Claude Fable 5.1 and Claude Mythos 5.1" and "Developing Enterprise Frontier Safeguards with our customers" (both 1 September 2026), the Claude Help Center articles "Covered Models" and "Data retention practices for Covered Models", CNBC's report of 1 September 2026, and OpenAI's "Offering Zero Data Retention for frontier models" (19 August 2026). Cost percentages are Anthropic's estimates. This briefing is general information, not legal or data-protection advice; confirm your own position with Anthropic and your DPO.