U.S. Commerce Lifts Export Controls on Claude Fable 5, and Anthropic Reintroduces Model with New Filter
The Department of Commerce lifted export controls after 18 days; Anthropic reactivates Fable 5 with a classifier that, according to the company, blocks over 99% of the jailbreak attempts described in the Amazon report.
On Tuesday, June 30, the U.S. Department of Commerce lifted export controls that had prevented Anthropic from offering the Claude Fable 5 and Mythos 5 models to foreign entities. The order was issued via a private letter on June 12, after Amazon researchers demonstrated the possibility of circumventing the locks on Fable 5 and obtaining, among other things, code to exploit software vulnerabilities. Anthropic began restoring access to Fable 5 globally on Wednesday, July 1.
During the 18 days that the ban was in effect, Anthropic suspended Fable 5 and Mythos 5 for its entire user base, both American and foreign, due to an inability to verify nationality in real-time. According to Paul Triolo, a partner at the consulting firm DGA-Albright Stonebridge Group, this is the first time a government has ordered a model developer to restrict access to a specific model based on nationality. The natural comparison to dual-use controls on semiconductors is complicated by one detail: hardware is traceable; a probabilistic model running in the cloud is not.
How Anthropic Returned to the Market
The agreement with Commerce included three concrete commitments: proactively detecting security risks in the models, working with the government on protocols for future releases, and reporting any observed malicious activity. The technical aspect involves the filter. Anthropic reintroduced Fable 5 with a cybersecurity classifier trained specifically against the vector described in the Amazon report. According to the company, the filter blocks over 99% of cases of that technique. When a request is blocked, it automatically reroutes to Claude Opus 4.8, which has reduced capabilities for offensive security tasks.
Fable 5 is returning first on the Claude Platform, at Claude.ai, Claude Code, and Claude Cowork. In public clouds, reintroduction on AWS, Google Cloud, and Microsoft Foundry will occur as each provider reprovisions. For Pro, Max, Team, and select Enterprise customers, the model will count towards quotas until July 7 as part of 50% of the weekly usage limit; after that, it will again consume credits.
What Washington Conceded, and What Beijing Gained
The 18-day window had repercussions on two fronts. First, Anthropic's international user base learned that access to a frontier model could be turned off by an administrative decision from the government, with immediate effect. Security teams in Frankfurt and London running Fable 5 for infrastructure code evaluations were left offline without any contractual clause supporting the suspension; some migrated to OpenAI's GPT-5.6 Sol, which is also under its own restrictions.
On the second front, the ban pushed developers outside the U.S. towards open Chinese models. Last week, the four most utilized models on OpenRouter came from DeepSeek, MiniMax, Tencent, and Xiaomi. The DeepSeek V4 Pro costs $3.48 per million output tokens; Fable 5 remains at $50 per million. The price gap explains much of the migration and likely influenced Commerce's calculations. For both Anthropic and the U.S. government, each day of suspension was a compilation advantage for Chinese laboratories.
Anthropic's response, in practice, was to absorb the regulatory requirement without abandoning the model. In addition to the classifier, the company claims it is drafting, with Amazon, Microsoft, Google, and other partners from the Glasswing consortium, a consensus framework on how to classify the severity of AI jailbreaks and how manufacturers should respond. If successful, it would represent the first formal attempt to standardize risk categories among laboratories that currently disclose incidents through their own channels.
An Unlocked IPO and an Open Precedent
Anthropic filed a confidential IPO request on June 1, with a valuation of about $350 billion. The offer was on hold while Fable 5 was offline, and the Commerce decision removed the primary short-term regulatory obstacle. What remains at risk is the precedent. The letter from June 12 did not undergo public consultation, did not cite detailed statutory authority, and was not discussed in Congress. Amazon researchers discovered a technique, Commerce reacted, and the company suffered 18 days without revenue from two models, while the global market adjusted suppliers. For the CIO in Tokyo, Berlin, or São Paulo who is now re-engaging Fable 5, the question is not whether the filter works; it is under what condition the next vulnerability report found by a partner could cause the service to shut down again.