US Fully Lifts Export Ban on Anthropic's Fable 5, Ending 18-Day Global Shutdown
The Commerce Department lifted its export controls on Anthropic's Fable 5 and Mythos 5 on July 1, restoring global access to Fable 5 and reinstating Mythos 5 for vetted US organizations after Anthropic agreed to new security safeguards.
The US Department of Commerce lifted its export control directive on Anthropic's Claude Fable 5 and Mythos 5 models on July 1, ending an 18-day shutdown that had taken both models offline worldwide. Anthropic said Fable 5 became available globally the same day across the Claude Platform, Claude.ai, Claude Code, and Claude Cowork, while Mythos 5 access was reinstated for a defined set of vetted US organizations.
How the ban unfolded
Commerce imposed the export control directive on June 12 after Amazon researchers reported a method for bypassing Fable 5's cybersecurity safeguards to elicit exploit code for a known vulnerability. Because Anthropic could not verify users' nationality in real time across its cloud distribution partners — AWS Bedrock, Google Cloud, Microsoft Foundry, Snowflake, and its own APIs — the company shut down access globally rather than risk noncompliance.
A partial reversal on June 26 restored Mythos 5 to more than 100 vetted companies and government agencies, but Fable 5, the more widely used consumer and developer model, stayed blocked. Anthropic said at the time it was continuing to negotiate a broader restoration.
Terms of the full restoration
Commerce Secretary Howard Lutnick said Anthropic no longer needs an export license for the models after agreeing to proactively detect and address security risks, cooperate with the government on shared standards for future model releases, and report malicious activity involving its models. Mythos 5's renewed access ties into Project Glasswing, a government cybersecurity initiative for a defined set of US organizations.
Anthropic said Fable 5 usage will count toward up to 50% of weekly limits for Pro, Max, Team, and select Enterprise plans through July 7, after which it transitions to a standard usage-credit model.
Part of a broader framework
The restoration follows OpenAI's negotiated preview of GPT-5.6 Sol to government-approved partners in late June, and both companies have since committed to pre-briefing federal officials ahead of future frontier model releases. Anthropic and other labs are also collaborating on an industry-wide jailbreak severity framework — an attempt to formalize how safety incidents like the one that triggered the Fable 5 ban get classified and disclosed before models reach the public.
Sources
AI-assisted reporting, overseen by the AgentsAI team. Spotted an error? Let us know.
Related agents
More ai news
Nvidia Backs OpenAI's Ohio Data Center With Up to $105 Billion in Financing
Nvidia will guarantee up to $105 billion in credit for a new SB Energy-built data center in Pike County, Ohio, that OpenAI will lease exclusively, deepening the circular financing ties between the chipmaker and its largest AI customer.
DeepSeek Finalizes Steep API Price Hikes, Ending Its Flat-Rate Era
DeepSeek confirmed a new peak/off-peak pricing structure for its V4 Pro and V4 Flash models effective August 16, raising some output-token rates by more than 1,100% as the low-cost Chinese lab moves away from the flat pricing that fueled its rise.
China's Z.ai Ships GLM-5.3, Claiming Coding and Cyber Gains Without a Bigger Base Model
Beijing-based Z.ai released GLM-5.3 on August 14, reusing its ~700B-parameter GLM-5.2 base model but claiming a 50% jump on internal coding benchmarks and a leading CyberGym score, positioning it against Anthropic and OpenAI on coding without training a larger model.
Researchers Show Encrypted Reasoning Traces Can Be Stolen Across OpenAI, Anthropic and Google APIs
A paper published August 10 by researchers from the ELLIS Institute Tübingen and the Max Planck Institute found that encrypted chain-of-thought blocks returned by OpenAI, Anthropic and Google reasoning APIs are interchangeable across sessions and models, letting a weaker model decode and leak a stronger model's hidden reasoning in plaintext.