Claude Fable 5.2 and zero data retention: what changes

Claude Fable 5.2 and zero data retention: what changes

Claude Fable 5.2 keeps prompts for 30 days: zero data retention becomes an exception. How OpenAI, Google, Copilot and Mistral compare, and why that matters.

Digital Sovereignty
9 min read

On November 9, 2026, Anthropic released Claude Fable 5.2, its most capable widely available model, 20% cheaper than the previous version. One line of the announcement matters directly to businesses: for most customers, prompts and responses will be retained for 30 days for safety monitoring. Zero data retention, the commitment by which a provider keeps no trace of requests, is granted only to "eligible" customers, and only for a transition period.

The rule is not new: it has applied since Fable 5, in June. But it unsettles companies for which zero data retention was a settled clause of their contract. With Fable, a clause in Anthropic's terms overrides any ZDR commitment already signed. Zero data retention is shifting from an option you switch on to an exception the provider grants.

Nothing forces anyone to give up Fable. A European company does, however, need to know what is kept, where, and under which law: data retention then becomes a sovereignty question, and OpenAI, Google, Microsoft and Mistral each answer it in their own way.

What is zero data retention?

Zero data retention (ZDR) is a contractual commitment: the AI provider processes the request, returns the response, then keeps nothing. Neither the prompt, nor the response, nor the files sent remain on its servers once the exchange is over.

It is not the default. At Anthropic, the API deletes inputs and outputs within 30 days, for every model, unless agreed otherwise. OpenAI keeps abuse monitoring logs for up to 30 days, Mistral for 30 rolling days, and Google's Gemini API for 55 days. These copies are used to detect abuse, with possible human review when content is flagged.

ZDR is negotiated. It takes prior approval from OpenAI, a justified request to Mistral, which "may approve or deny it", and an enterprise agreement with Anthropic. Anthropic itself acknowledges that many companies, "especially in regulated industries", find it hard to use models that retain data. For them, ZDR is the guarantee that a confidential piece of data sent to a model exists nowhere else once the answer comes back.

ZDR also has gaps, even when granted. At OpenAI, prompt caching, code execution containers and MCP servers leave temporary traces. At Mistral, it does not cover stateful APIs: agents, batch jobs, files, conversations. A ZDR contract only protects the calls it covers, and the list varies from one provider to the next.

What does Claude Fable 5.2 change for your data?

Anthropic classifies its most powerful models as "Covered Models", models that cross capability thresholds warranting additional safeguards. Fable 5 and Mythos 5 were designated on June 9, 2026, Fable 5.1 on August 31, Fable 5.2 on November 9. For these models, prompts and responses are retained for at least 30 days and then deleted, unless subject to a safety investigation or a legal obligation.

The API documentation spells out the consequence: ZDR "is not available" for the six Fable and Mythos models, "unless expressly authorized by Anthropic". An organization configured for ZDR that calls Fable gets an error: "your organization or workspace must have data retention enabled". Anthropic's service-specific terms state that this retention "supersedes any modified retention commitments (e.g., ZDR)". A ZDR agreement signed last year does not cover Fable.

The transitional exception and Enterprise Frontier Safeguards

Anthropic has left a door open. "Eligible" customers can use Fable 5, 5.1 and 5.2 with ZDR for their internal applications, for a limited time. On Bedrock, AWS gives a date: through at least June 30, 2027. Anthropic does not publish its eligibility criteria and reserves the right to withdraw the arrangement.

This period is meant to lead to Enterprise Frontier Safeguards (EFS), announced on September 1. The idea: logs stay in the customer's cloud (Amazon S3, Azure Blob Storage, Google Cloud Storage), Anthropic's automated monitoring runs there, and alerts go to the customer, with no review by Anthropic employees. Anthropic says the service is free. With Fable 5.2, it moved into private beta with Bridgewater, Salesforce and Stripe. For every other customer, it means 30 days of retention at the provider.

Why Anthropic retains data, and what it promises

The rationale is serious. Some attacks only become visible across multiple requests: a "best-of-N jailbreak" sends hundreds of variants of the same request until one slips past the filter. Detecting it requires keeping the history. And the fear of offensive use is not theoretical: Claude Mythos, the same underlying model as Fable without some of its safeguards, finds vulnerabilities faster than they can be patched.

Anthropic commits not to use this data to train its models, nor for any purpose other than safety. By default, no Anthropic employee can read retained conversations; human review only happens on flagged content, by a small group of people, and every access is logged. Flagged content, however, can be kept for up to two years.

These safeguards are real, but they are declarations. A software vendor summed it up on the Hacker News thread about the rule (609 points in June): "due to this policy, I cannot and will not use Fable in the products we sell". For a SaaS vendor, Anthropic becomes a subprocessor that keeps its own customers' data, and those customers have to be told. Forrester advised as early as June moving vendors that use Fable into a higher risk tier.

Who keeps the data, and where?

The rule "follows the model": it applies wherever Fable is offered, including on the big clouds. Going through AWS, Google or Microsoft does not get around it, but it changes who keeps the data and where.

ChannelWho retainsLocation for a customer in France
Anthropic APIAnthropic"global" or "us", no Europe option
Amazon BedrockAWS, up to 30 days"Global" profile only from Paris or Frankfurt
Microsoft FoundryAnthropicAnthropic's infrastructure, possibly outside the Azure geography
Google CloudGoogle, with sharing to AnthropicEU multi-region available

At Anthropic directly, the location parameter only accepts "global" or "us": no guaranteed inference in Europe. On Bedrock, AWS retains all Fable traffic for up to 30 days and its staff may review flagged requests; the Fable 5.2 model card only offers the "Global" profile from European regions, and retained data is stored in the destination region. On Foundry, Fable 5.2 runs on Anthropic's infrastructure, and data "might be processed outside of your selected Azure geography".

Google Cloud is the only channel listing an EU multi-region for Fable 5.2. The same page, however, asks customers to enable "prompt-response sharing with Anthropic", while Anthropic writes that retained data stays with the cloud provider. We found no source that resolves this contradiction.

OpenAI, Google, Microsoft, Mistral: who still offers zero data retention?

Anthropic is not alone in monitoring its frontier models. The four other major providers have made different, sometimes opposite, choices.

ProviderDefault retentionZero data retentionFrontier modelProcessing in Europe
Anthropic (API)30 daysBy agreement, excluded for Fable unless exempted30 days required, EFS in betaNo
OpenAI (API)30 daysOn approvalZDR kept, logs can stay with the customerYes, with ZDR or a similar regime, +10%
Google (Vertex AI)On flagged activity, up to 90 daysYesGemini 4 Carbon: 30 days in the customer's environmentYes, EU region or multi-region
Microsoft 365 CopilotLike other Microsoft 365 contentNot applicableFable outside Microsoft's DPA, off by defaultEU Data Boundary, Claude excluded until end of 2026
Mistral (API)30 daysOn request, stateless callsLabs models excludedEU by default

OpenAI makes the opposite bet

In August 2026, OpenAI published a post titled "Offering Zero Data Retention for frontier models", which The Next Web read as a response to providers that require retention: ZDR remains available on its frontier models. To reconcile safety and confidentiality, OpenAI offers Private Safety Processing: encrypted logs, kept for 30 days, are stored by the customer, and OpenAI only keeps "an index with operational metadata". It is the same principle as EFS.

Two caveats. OpenAI keeps a clause allowing it to make a model ineligible for ZDR for a given customer, after written notice, in case of severe risk. Its Europe region, for its part, requires ZDR or a similar regime, with a 10% surcharge on recent models.

Google keeps logs in the customer's environment

On Vertex AI, Google applies to its standard models monitoring limited to flagged activity, from which customers under a master agreement are exempt. For its most advanced models (Gemini 4 Carbon and its cyber variants), Google logs prompts and responses for up to 30 days, but "inside the customer's environment"; for customers under a master agreement, Google considers these logs ZDR-compliant. Retention exists at Google too, but the stock stays with the customer.

The consumer-facing Gemini API (AI Studio) is a different world: 55 days of retention with no exemption, and Google itself points to Vertex for anyone who needs real ZDR.

Microsoft takes Fable out of its contract

Microsoft 365 Copilot does not run abuse monitoring: its services have opted out of it. Conversations are stored like other Microsoft 365 content, accessible to administrators and kept until a policy deletes them.

Fable is handled separately. Microsoft puts Fable 5 and Mythos 5 in an "Anthropic models with Data Retention" category: these models are off by default, Anthropic acts as an independent processor, and the data falls outside Microsoft's contract and DPA. "Anthropic (not Microsoft) stores most inputs and outputs for up to 30 days." All Anthropic models are also excluded from the EU Data Boundary, the perimeter within which Microsoft keeps European data processing. Microsoft plans to bring them in by the end of 2026, without saying whether Fable-class models are included.

Mistral: European, but not retention-free

Mistral keeps 30 rolling days by default, like the others. Its ZDR is limited to paid plans and stateless calls, at its discretion. For its Labs and Preview models, privacy preferences, ZDR included, do not apply: data sent to these models can be used for training. Data is hosted in the European Union by default and the company is French, but part of its compute runs at US subprocessors, as we noted when looking at who owns France's tech champions.

Why retention is a sovereignty issue

Thirty days of retention create a stock. And a stock held by a US company falls under the CLOUD Act, which requires a provider to "preserve, backup, or disclose" data under its control, "regardless of whether such communication, record, or other information is located within or outside of the United States". Storing data in Europe changes nothing when the parent company is American. Questioned by the French Senate in June 2025, Microsoft France's director of public and legal affairs could not guarantee that French customers' data would never be handed over to US authorities.

ZDR reduced this risk to almost nothing: no one can demand data that no longer exists. Without ZDR, the provider's seriousness matters less than the law that applies to what it keeps. Customer-side storage models (Anthropic's EFS, OpenAI's Private Safety Processing, Gemini logs at Google) move the stock to the company, and therefore under its own jurisdiction. That is a real improvement, with a trade-off: according to a Gartner analyst quoted by CSO Online, the customer inherits alert triage and incident response.

The legal framework for transfers remains fragile. After a US Supreme Court ruling of June 29, 2026 on the independence of federal agencies, the European Data Protection Board asked the European Commission to assess whether the Data Privacy Framework, the agreement that allows these transfers to the United States, still works. Anthropic, for its part, relies on standard contractual clauses, not on that framework.

The only way to remove the question is to choose a European provider controlled in Europe. Scaleway applies zero data retention by default on its AI APIs, hosted in Europe, and states that it is not subject to the CLOUD Act; OVHcloud also advertises zero data retention on AI Endpoints, hosted in the European Union. The limitation is well known: they serve open-weight models, not Fable. We have covered elsewhere what dependence on US providers means for a European company.

What an IT department can take from this

Anthropic itself recommends starting most work with Opus 5.5, which keeps the negotiated retention regime. Fable is reserved for the most ambitious work, and the remaining question is which data can go to a model that keeps it for 30 days. We would approach it in four steps.

Contracts first

A ZDR commitment signed with Anthropic, AWS or Microsoft does not cover Fable, unless there is a written exemption. Anthropic lets you isolate Fable in a workspace with retention enabled, while other workspaces stay on ZDR. That is the minimum before opening Fable to teams.

A boundary for each type of data

An audit of public code or a document watch can live with 30 days of retention. An M&A file, health data or industrial secrets cannot. The boundary is drawn use case by use case, not provider by provider.

The actual path of the data

From Paris, Fable on Bedrock goes through a global profile, and on Foundry through Anthropic's infrastructure. The GDPR record of processing activities and the data protection impact assessment must mention a subprocessor that keeps personal data for 30 days, and the country where it does so.

Routing by sensitivity

An LLM gateway can send each request to the right model: Fable for hard tasks on ordinary data, an open-weight model hosted in Europe for sensitive data. That is the choice we made at Ask This Guy, which does not use Anthropic: each customer organization picks an "EU only" or "worldwide" AI policy, and the first sends nothing to a US provider. Our sovereign AI page details these criteria.

Frequently asked questions about zero data retention and Claude Fable 5.2

Is Claude Fable 5.2 available with zero data retention?

Not by default. Anthropic requires 30-day retention of prompts and responses for safety monitoring, and this rule overrides existing ZDR agreements. Only enterprise customers deemed eligible can use it with ZDR for their internal applications, on a transitional basis, until Enterprise Frontier Safeguards is ready. On Amazon Bedrock, this arrangement runs through at least June 30, 2027.

Does Anthropic use retained data to train Claude?

No, according to Anthropic. Fable data is retained for safety purposes only, deleted after 30 days unless flagged or legally required, and no employee can read it by default. Human review only happens on flagged content, and every access is logged. These are contractual commitments and declarations.

Is Claude GDPR compliant?

Anthropic offers a data processing addendum that covers transfers to the United States through standard contractual clauses. But Anthropic offers no guaranteed inference in Europe and, for Fable, keeps data for 30 days. Compliance therefore depends on the use case: legal basis, information to the people concerned, impact assessment and the nature of the data sent. For sensitive data, a European provider with zero data retention by default significantly reduces the risk.

Does zero data retention protect against the CLOUD Act?

Largely, because data deleted after processing cannot be demanded. But ZDR does not cover everything: cache, files, agents or safety classifier results may remain stored. And during processing, the data does pass through the infrastructure of a provider subject to US law. A European provider controlled in Europe is not subject to it.

Which AI providers still offer zero data retention?

OpenAI offers it on approval, including on its frontier models, and requires it for its Europe region. Google offers it on Vertex AI, and stores the logs of its most advanced models in the customer's environment. Mistral grants it on request for stateless calls. Scaleway and OVHcloud apply it by default on their open-weight models hosted in Europe. Anthropic excludes it for Fable, unless exempted.


Fable 5.2 may be setting a norm: the more powerful a model, the more its provider wants to watch how it is used. Anthropic imposes it, Google and OpenAI negotiate it by leaving the logs with the customer. For a European company, knowing whether the provider keeps its data is no longer enough: it needs to know who could one day ask the provider to hand it over. Better to check before plugging in the next frontier model.

Partager :

Related Articles

Gemini for Healthcare and France's reliance on US techDigital Sovereignty

Gemini for Healthcare and France's reliance on US tech

Gemini for Healthcare targets US hospitals. In France, the Health Data Hub already sits on Microsoft Azure and Doctolib's AI uses Google and Anthropic models.

11 min
CoreWeave, Mistral AI's subcontractor: should you worry?Digital Sovereignty

CoreWeave, Mistral AI's subcontractor: should you worry?

For all its products, Mistral AI declares just two compute operators: its own servers and CoreWeave, whose debt is in distress. What a default would change.

Best French AI for enterprises in 2026: 14 vendors put to the sovereignty testDigital Sovereignty

Best French AI for enterprises in 2026: 14 vendors put to the sovereignty test

Fourteen French enterprise AI vendors checked against public sources: who owns the capital, who hosts your data, who runs inference.

12 min

Interested in our solutions?

Discover how Ask This Guy can help you accelerate with AI

Book a Demo