Fable 5 is again available on multiple Claude endpoints under tightened access rules and operational restrictions imposed after a jailbreak report.

Availability and main changes

Fable 5 can now be deployed on Claude Platform, Claude.ai, Claude Code and Claude Cowork, with temporary free access until 07.07.

The model now embeds a classifier that blocks a specific jailbreak technique in 99% of detected attempts, and routine coding tasks are routed to Opus 4.8 instead of Fable.

Mythos 5 remains unavailable to the public and continues to be used only by selected U.S. companies.

Government agreement and oversight

Anthropic co‑founder Tom Brown signed an agreement with U.S. authorities establishing mandatory pre‑release testing of powerful models by government agencies.

• The lab must hand over all future high‑capability models for independent safety testing before release.
• Anthropic is required to report detected malicious user activity to the government promptly.
• A consortium with Amazon, Google and Microsoft will provide a unified scale for assessing jailbreak severity.

Sam Altman has reportedly proposed transferring 5% of OpenAI ownership to government stakeholders, according to discussions mentioned in the report.

Operational recommendations for developers

Practitioners are advised to use Fable 5 as an orchestrator and delegate heavy or routine workloads to specialised subagents.

A suggested setup places Fable 5 as the primary reasoning model, Opus for architecture and complex debugging, and Sonnet for mechanical routines such as tests and formatting.

Integrating a Codex‑style plugin can provide an alternative engineering opinion and parallel validation of outputs during development workflows.

Detection and regional restrictions

Claude Code now detects attempts at model distillation originating from certain jurisdictions, and inserts invisible markers into system prompts to hinder unauthorized distillation.

These measures aim to reduce jailbreak vectors while preserving limited developer access for permitted use cases and security testing.