On June 9, Anthropic released Claude Fable 5, the first publicly available model in the Mythos family, capable of handling complex, multi-stage tasks spanning several days. Microsoft launched the model on GitHub Copilot and Azure Foundry the same day, yet its internal employees were unable to access it. The controversy centers on Fable 5’s requirement to retain data for 30 days for security monitoring, conflicting with Microsoft’s long-standing zero-data-retention policy designed to protect customer code and confidential information. Microsoft’s legal team is currently evaluating Anthropic’s new data retention requirements, with outcomes still uncertain. For financial, healthcare, and government clients, ZDR is often explicitly written into contracts and represents a hard line for Microsoft. Anthropic, however, states that data retention is a prerequisite for opening up Mythos-level capabilities, aimed at detecting large-scale abuse and attacks.

Article author and source: AI New Era

On June 9, Anthropic officially released Claude 3.5.

This is the first Mythos family model designed for the public, claiming to handle complex, multi-stage tasks lasting several days.

Immediately after the release of Fable 5, Microsoft rolled it out to customers; GitHub Copilot and Microsoft Foundry were available the same day, and the Azure official blog headline read: “Unlocking the Next Era of Autonomous Agents.”

Claude Fable 5 is now selectable in the GitHub Copilot model selector, but this option is not available in the internal version for Microsoft employees.

Microsoft Azure's official blog proudly announced the launch of Fable 5 on Foundry: "Opening the next era of autonomous agents."

Customers get to try it first, but our own employees can't use it.

According to The Verge, among the Claude models available within GitHub Copilot, used internally by Microsoft employees, all other Claude models are present except for Claude Fable 5.

Reports indicate that Microsoft has been telling employees that its legal team is evaluating Anthropic’s new data retention requirements, with primary concerns centered on customer data and confidential information.

No one knows for sure whether legal will ultimately approve it.

A red line blocked the strongest Claude.

To understand what Microsoft is concerned about, first look at the data retention policy on Anthropic's official Fable 5 model release page, which states:

To use Fable, 30 days of data retention is required for security monitoring.

On its changelog, GitHub conveyed this upstream term to Copilot users:

Unlike other Claude models in Copilot, Claude Fable 5 requires data retention to run Anthropic's safety classifier. Prompts and outputs are retained by Anthropic for up to 30 days, after which they are deleted.

All other Claude models, including Opus 4.8, Sonnet 4.5, and Haiku 4.5, continue to operate under the Zero Data Retention (ZDR) policy.

Anthropic's Privacy Center states that inputs and outputs flagged by the safety classifier as violating usage policies may be retained for up to two years.

This means that 30 days is merely the standard upper limit: once the classifier triggers a flag, the period during which enterprise data is out of the company’s control extends from one month to two years.

Retained data will not be used to train models; it is solely used for security monitoring. This is a commitment stated in Anthropic’s terms, and GitHub has also confirmed this in its changelog.

But for Microsoft itself, the issue is not here.

What it holds are code from developers worldwide on GitHub, contracts within Office documents, and corporate secrets on Azure cloud.

Each prompt from an employee may contain customer data. Storing this information on third-party servers for 30 days is inherently a seed for compliance violations.

The logic of ZDR is simple: once a request is processed, the data is immediately deleted. With no copies stored on the server, there can be no leakage, no retrieval, and no unauthorized access.

For clients in finance, healthcare, and government, ZDR is often directly included in contracts. It is Microsoft’s commitment to its customers and also the line it has drawn for itself.

Meanwhile, Fable 5's new policy on 30-day retention data directly intersects with this line.

Anthropic: We won't release the model without reviewing the data.

Turning to the other side, Anthropic’s reasoning is equally strong: trust and security.

To ensure the responsible deployment of Mythos-level models, we require limited data retention and review in our security practices. On all platforms offering these models, prompts submitted to Mythos-level models and their generated outputs are retained for 30 days.

This policy took effect on June 9, 2026, and applies to the Mythos series models as well as future models with similar capabilities that Anthropic officially designates as covered models.

For all other models, the terms remain unchanged, and users' usage content is unaffected.

Several months ago, Anthropic stated that models like Mythos are too powerful for cybersecurity tasks and too dangerous to be publicly released.

Claude Fable 5 and Claude Mythos 5 share the same underlying model. That means Fable 5 also possesses the same capabilities as the brain, deemed too dangerous to release publicly, just a few weeks ago.

The difference is that Fable 5 adds an additional security layer: requests involving sensitive areas such as cybersecurity and biology are intercepted by a security classifier and automatically routed to Claude Opus 4.8, without being charged at Fable’s rate.

Truly unrestricted Mythos 5, available only to a select few approved organizations.

So why keep the data at all? Anthropic's Help Center explains that some attacks cannot be detected by looking at a single request alone.

Best-of-N jailbreaking involves sending hundreds of finely tuned prompt variations to the model, betting that one will bypass its defenses—such as in large-scale abuses like data extortion. Only by examining vast numbers of requests together do patterns emerge.

To detect these threats, prompts and outputs must be temporarily stored and analyzed together.

Anthropic also listed protective measures:

Employees are not permitted to access retained data at will. Only a limited number of approved reviewers may access such data using a dedicated tool, and only when the content is flagged as potentially severely harmful or upon a customer’s written request. This tool prohibits exporting, copying, or downloading.

Each visit is recorded in an immutable log. Data is automatically deleted after 30 days.

Eligible organizations can also add customer-managed encryption keys and access transparency audit logs.

According to Anthropic's explanation, it's not that it wants to see your data; rather, if it doesn't see it, it dares not release the model.

Thirty-day retention is a requirement to unlock Mythos-level capabilities. To use them, you must first relinquish your ZDR红线.

Who is affected, and how to set up retention?

Individual users are not affected.

Claude’s Free, Pro, and Max plans have always retained data for security purposes, and nothing has changed with this update.

What has truly changed are the organizations that have specifically configured ZDR:

• Set up the ZDR workspace in Claude Console;
• Running Claude Code on ZDR in Claude Enterprise;
• And enabled with ZDR via AWS Bedrock, Google Cloud Agent Platform, or Microsoft Foundry.

Companies that initially had stricter data compliance requirements face the greatest challenge in making new choices.

The official guide for using Fable 5 has been provided.

Anthropic's Help Center lists retention configuration methods by platform; among the five channels, only Azure Foundry (highlighted) requires a separate subscription.

For those using the Claude API directly, enable retention in the developer console for the specified workspace by navigating to Workspace → Management → Privacy Controls; all other ZDR workspaces remain unchanged.

Claude Code follows its workspace—if the workspace is open and active, it can use the designated model.

For Amazon Bedrock and Google Cloud Agent Platform, you must also enable retention first, but there is a key difference: the retained data stays within your own AWS or GCP environment and never leaves your cloud.

For companies that prioritize data ownership, this is a crucial buffer.

Azure Foundry is unique: retention is configured per subscription; if a subscription has ZDR enabled and you want to use Fable 5, you must create a separate one.

For the ZDR organization using Claude Enterprise, the management console toggle is now live, and the primary administrator can modify it directly.

If you don't want to touch the production environment, Anthropic can help set up a separate sandbox organization.

The design concept is isolation: the only access point for retention is within the workspace, subscription, or sandbox you specify; everywhere else, the red line remains unchanged.

You can do it, but you must sign it yourself.

The story isn't over yet—Microsoft also held back something from its own customers.

A subtle detail in the GitHub changelog: In Copilot Business and Enterprise, the Claude Fable 5 policy is disabled by default and must be manually enabled by administrators.

Enabling this strategy constitutes confirmation of the data retention requirements.

GitHub officially states: the strategy is disabled by default; enabling it constitutes confirmation that you accept the retention requirements. This toggle is the "Disclaimer" mentioned in this section.

Similarly for Azure.

To use Fable 5 with a ZDR subscription, a separate subscription must be created; initiating a new subscription is a choice made by the customer themselves.

In other words, Microsoft bundles the choice and responsibility together and hands them over to the customer’s administrator. If you want to use it, you sign off on it yourself.

That default-off switch is like a disclaimer from Microsoft.

It follows the same logic as restrictions on employees; the only difference is that customers can still choose for themselves, while company employees have no choice.

Choosing a model? Don't just look at benchmark scores.

Previously, to judge whether a model was usable, we looked at benchmark scores, price, and user experience.

There is now an additional, invisible hurdle—legal compliance.

Which AI you can use within your company depends not only on cost-effectiveness but also on legal issues such as data compliance; Fable 5 is a case in point—even Microsoft’s own employees were blocked from accessing it, let alone outsiders.

Anthropic, of course, hopes to release stronger models earlier than its competitors, but the more powerful the model, the higher the safety safeguards must be.

It bets that Mythos-level capabilities are worth users surrendering their 30-day data retention rights.

On Microsoft’s side, the legal team is evaluating whether the ZDR红线 is worth making an exception for Fable 5.

Microsoft's internal evaluation has not yet reached a conclusion.

One thing has changed: going forward, to access the most powerful models, you may need to pass legal review first.