Anthropic Redeploys Claude Fable 5 Globally With New Safeguards and Usage Limits

Anthropic has redeployed Claude Fable 5 to users globally from 1 July 2026, following a brief pause to address a security vulnerability. The redeployment comes with a new safety classifier, revised usage limits for the first week of access, and confirmed US government approval for a separate advanced model, Claude Mythos 5.

The full details are published in Anthropic’s official announcement at anthropic.com

What Prompted the Pause

The temporary withdrawal of Fable 5 followed a vulnerability report submitted by Amazon. The reported issue concerned a bypass of the model’s safety guardrails, a category of problem that AI developers treat with significant caution given the potential for misuse at scale.

Anthropic’s response was to develop a new safety classifier specifically targeting the behaviour described in the report. According to Anthropic’s announcement, the classifier is designed to catch 99% of bypass attempts, representing a meaningful improvement over what was in place at the time of the vulnerability report.

The development of that classifier, combined with coordination with the US government, accounts for the delay between the original planned deployment and the 1 July redeployment date.

The New Safety Classifier

A classifier in this context is an automated security layer that evaluates each user prompt before it reaches the model. Its role is to determine whether a request is a standard query or an attempt to circumvent the model’s safety guidelines for harmful purposes.

Anthropic has been direct about one limitation of the new classifier: it may generate false positives, particularly during routine coding and debugging tasks. Users whose requests are blocked will receive a notification and their request will be automatically rerouted to Claude Opus 4.8, keeping workflows functional even when Fable 5 declines to process a specific input.

Anthropic has committed to improving the classifier over time to reduce the rate of false positive incidents.

Usage Limits and Access Timeline

Access to Fable 5 from 1 July through 7 July is subject to up to 50% of weekly usage limits for Pro, Max, Team, and select Enterprise plans. From 8 July, access moves to usage credit pricing.

Anthropic’s announcement sets out the full picture: Fable 5 is available on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. Restoration of access on AWS, Google Cloud, and Microsoft Foundry is in progress, with Anthropic stating it will move as quickly as possible on those integrations.

Mythos 5 and Government Coordination

Separately, Anthropic has restored access to Claude Mythos 5 for a set of US organisations, following approval from the US government on 26 June 2026. US Secretary of Commerce Howard Lutnick confirmed the approval publicly, noting that the government had worked closely with Anthropic over the preceding two weeks to analyse and approve Fable 5 and ensure alignment with US government requirements.

Mythos 5 access remains limited to a defined set of US organisations at this stage. Anthropic has stated it is continuing to coordinate with the government to expand access to a broader set of domestic and international partners through its Glasswing programme.

What This Means for Teams Using Claude

For teams already using Claude across their workflows, the practical impact of the redeployment is limited. The rerouting mechanism means that a blocked Fable 5 request reaches Opus 4.8 automatically rather than failing outright, which preserves continuity for most use cases.

The usage limit through 7 July is the most immediate consideration for heavy users. Teams operating at or near their weekly usage ceiling will need to factor in the 50% cap when planning workloads in the first week of access.

The false positive risk during coding and debugging tasks is worth noting for development teams. The rerouting to Opus 4.8 provides a fallback, but teams running high-volume or time-sensitive engineering workflows may want to monitor for unexpected blocks during the period before Anthropic’s classifier improvements are implemented.