Claude Fable 5, Anthropic’s most capable model, is now accessible to subscribers across Max, Pro, and Team plans following the removal of a U.S. Department of Commerce export restriction. The rollout, however, has been met with frustration from early users who say the restored model behaves noticeably differently from its pre-ban incarnation.

Availability Comes With Significant Caveats

While inclusion in paid plans is a positive signal, Anthropic has imposed hard caps on Fable 5 usage. Subscribers are limited to spending no more than 50 percent of their weekly usage allowance on the model. That constraint is scheduled to tighten further after July 7, when Fable 5 transitions to a strictly usage-credit model, requiring direct payment beyond any subscription tier.

Safety Guardrails Are Triggering False Positives

The more immediate concern for developers is behavioral, not financial. Users on Reddit and in developer forums report that Fable 5 frequently falls back to Opus 4.8 mid-session, often on tasks that carry no obvious safety risk. Claude apparently notifies users when this switch occurs, making the fallback visible and disruptive to workflow.

Reported triggers include:

  • Searching a codebase for dead or unused code
  • Systems-level work in C, C++, and Rust
  • References to Win32 API or memory management
  • Project files or prompts containing words such as “security,” “vulnerable,” “unsafe,” or “hook”

One developer described the model as unusable for certain low-level coding tasks, while another noted it was “very very obvious” when the fallback triggered because Claude announces the switch explicitly.

The Model Itself Has Not Been Modified

Importantly, the underlying Fable 5 model weights do not appear to have been changed. The degraded experience seems to stem from Anthropic applying a wider safety margin in its guardrail systems, which are intercepting and rerouting requests more aggressively than before. This distinction matters: when Fable 5 actually processes a task, its output quality appears consistent with prior assessments. The problem is that the safety layer is catching a disproportionate number of legitimate requests.

Anthropic has publicly noted that its updated safeguards rely on a broad safety margin, a design choice that likely explains the high false-positive rate users are experiencing. The company has not yet formally acknowledged the false-positive reports, though the volume of user feedback makes it probable the issue is on Anthropic’s radar. A guardrail tuning update would be the expected remediation path.