Claude Fable 5: The Model That Got Banned, Then Came Back

Because Anthropic could not reliably distinguish foreign nationals from domestic users in real time, it disabled both models for all customers worldwide on June 13 .

What Triggered the Ban?

The immediate cause was a reported jailbreak by researchers at Amazon and several other companies, who allegedly extracted information useful for conducting offensive cybersecurity operations — what the government viewed as a national security concern . According to Anthropic's internal review, the prompt that triggered the export control order was a standard codebase security audit request — a capability Anthropic argued already exists in OpenAI's GPT-5.5 and other frontier models .

The 18-Day Suspension

The export controls were lifted on June 30, 2026, ending an 18-day suspension . Fable 5 was restored to users on July 1 across Claude.ai, the Claude Platform API, and enterprise plans . The Commerce Department's letter reserved the right to reimpose controls if Anthropic's safety commitments are not upheld .

The Bigger Story: A Jailbreak Severity Framework

The Fable 5 saga did not just produce a restored model — it forced the creation of a new industry-wide framework for scoring AI jailbreak severity. Anthropic announced it is working with Amazon, Microsoft, and Google (along with other Project Glasswing partners) on a shared rubric . Key characteristics:

• It scores how far the bypass went, what capabilities became reachable, how repeatable the attack was, and whether the exposed behavior had real operational consequences .
• The White House was also involved in negotiating technical standards to grade jailbreak severity as part of the deal to lift controls .
• The framework is intended to give developers and regulators a shared rubric for determining how serious a jailbreak is and what response is warranted, acknowledging that no model can be made fully jailbreak-proof .

This framework may turn out to be more consequential than the model itself. As one analysis put it: "Anthropic, Amazon, Microsoft, and Google are developing shared jailbreak-severity standards, creating a de facto industry policy in the absence of federal AI regulations" .

How Fable 5's Safety Classifiers Work

Fable 5 launched with a new architecture: safety classifiers that flag certain high-risk request categories and route flagged queries to Claude Opus 4.8 (a fallback model) instead of letting Fable 5's full capabilities respond . The flagged categories include:

• Cybersecurity / offensive cyber operations — requests related to vulnerability identification, codebase ingestion for offensive purposes, or other cyberattack-enabling queries
• Biology and chemistry — pathogen characteristics, synthesis or cultivation of microorganisms, gain-of-function research, biological toxins, lab protocols applicable to select agents
• Distillation-risk prompts — attempts to extract or reproduce the model's weights, training data, or proprietary behavior

After the export controls were lifted, Anthropic deployed an improved safety classifier specifically targeting the jailbreak technique that triggered the original order, claiming it blocks the reported technique in over 99% of cases, with blocked requests still routed to Opus 4.8 . The company acknowledged the classifier will flag some benign requests as a tradeoff .

Fable 5 Pricing: What It Costs

API pricing: Claude Fable 5 is priced at $10 per million input tokens and $50 per million output tokens on the Anthropic API . The official Claude docs also list $12.50/MTok for cached input, $20/MTok for non-cached input, and $1/MTok for "bookend input" .

Free introductory period: From launch on June 9 through June 22, 2026, Fable 5 was included at no extra cost on Claude Pro, Max, Team, and seat-based Enterprise plans .

Transition to metered billing: On June 23, 2026, Fable 5 was removed from those bundled plans. After that date, any Fable 5 request draws down paid usage credits billed at the API per-token rates ($10/$50 per MTok) rather than being covered by the plan's included allowance .

A note on July 7: Some third-party commentary mentions a July 7 date for a subsequent pricing adjustment, but the official Anthropic documentation and all detailed third-party reports agree on the June 23 cutoff for the free-inclusion period . After the export controls were lifted on June 30, some users on Pro, Max, and Team plans saw a temporary re-inclusion window through July 7 before the metered credits took full effect .

The Takeaway

Claude Fable 5's 18-day journey from launch to ban to restoration is not just a story about one model. It is a case study in how frontier AI companies and governments are — awkwardly, publicly — negotiating the rules for deploying powerful models in a world that lacks formal AI safety regulation. The jailbreak severity framework that emerged from this standoff may set the standard for how the industry responds to the next inevitable bypass.