TL;DR
A June 12 U.S. export-control directive forced Anthropic to disable Fable 5 and Mythos 5 worldwide, according to Anthropic statements and reporting cited by Thorsten Meyer AI. Weeks earlier, OpenAI completed its retirement of GPT-4o from ChatGPT. The two cases show that many AI users rent access to models they do not control.
On June 12, the U.S. government issued an export-control directive that forced Anthropic to disable its Fable 5 and Mythos 5 models worldwide, according to Anthropic statements and reporting cited by Thorsten Meyer AI. The shutdown, paired with OpenAI’s February retirement of GPT-4o, shows that many people and businesses depend on AI systems they can access but do not control.
The directive was aimed at foreign-national access to Anthropic’s newest models, including foreign nationals inside the United States and Anthropic’s own employees, according to the source material. Because Anthropic could not limit access narrowly enough to meet the order, the company disabled both models globally on short notice.
Anthropic’s account says the letter arrived in the evening, gave no detailed rationale, and left the company about 90 minutes to comply before the models went offline by midnight. The stated basis was national security, but the specific risk finding behind the order has not been made public.
OpenAI’s GPT-4o case was different. OpenAI announced the retirement of GPT-4o, GPT-4.1, GPT-4.1 mini and o4-mini in late January and removed them from ChatGPT on February 13. API support was also scheduled for removal, meaning applications with old model identifiers would eventually receive errors unless they migrated.
The Switch: You Never Owned It
In 2026 a government turned off a frontier model worldwide in ~90 minutes — and a company retired a beloved one with ~2 weeks’ notice. You don’t own the model you build on. You access it. Access can be revoked.
Access is the only chokepoint that flips in an afternoon — and the version that hits you won’t be Washington, it’ll be a deprecation. Open weights you host can’t be deprecated, geofenced, repriced, or revoked. Short of that: route through a provider-agnostic gateway, keep a tested fallback, and treat every model string as a dependency that will be pulled.
Hosted Models Can Vanish Quickly
For companies that use hosted AI through APIs, the issue is operational control. A model may power customer service, coding workflows, compliance reviews or internal search, but the provider or a government order can change availability faster than many production systems can adapt.
The business risk is not limited to shutdowns. Access can be changed through regional blocks, pricing moves, rate limits, terms restrictions or model behavior updates. Those actions can alter costs, break workflows or force urgent engineering work even when the provider has sound reasons for retiring older systems.
personal AI model hosting device
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Two Switches, One Dependency
Thorsten Meyer AI’s Control Series frames model access as a sharper chokepoint than compute, data or power because it can be pulled in a single decision. Export controls were designed around controlled goods and infrastructure, but hosted AI turns access into a software permission problem.
The Anthropic order shows the state version of that control. The GPT-4o retirement shows the product version: OpenAI had previously faced backlash after removing older models in 2025, later restoring access for some users, but its 2026 retirement plan went ahead after OpenAI said use of GPT-4o had fallen to a fraction of a percent, according to reporting cited in the Control Series report.
“no detailed rationale”
— Anthropic’s account cited in the Control Series report
Personal AI Servers: A Guide to Building Private AI Infrastructure for Secure, Offline and Self-Hosted Local LLMs for Data Privacy
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Security Rationale Still Disputed
The public record still leaves gaps. The government has not released the detailed security analysis behind the Anthropic directive, and it is unclear whether any narrower compliance method could have kept access available for some users.
It is also unclear how many Anthropic customers were affected, how long the controls will last, and whether future export rules will target model access more often. For OpenAI customers, the open question is less legal and more practical: whether migration plans can keep pace with provider schedules.
AI Systems in Production: Operating, Controlling, and Governing Intelligent Systems After Launch (Engineering Intelligence: Designing, Operating, and Governing AI Systems)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Migration Plans Face Tests
Anthropic is expected to keep pressing U.S. officials for a path to restore access or define narrower limits. Any public settlement, safety review or revised export-control language would set a reference point for other frontier-model providers.
Developers and companies are likely to respond by treating model identifiers as unstable dependencies. The near-term steps are audits of hardcoded model strings, provider-agnostic routing, tested fallback models and, where workloads justify it, self-hosted open-weight models.
Small Language Models (SLMs): The Complete Practical Guide to Building Fast, Local, and Private AI Systems: A Step-by-Step Professional Handbook for Developers, Entrepreneurs, and AI Builders
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Did Anthropic shut down Fable 5 and Mythos 5 on its own?
No. According to the source material, the June 12 directive restricted access by foreign nationals and left Anthropic unable to comply except by disabling Fable 5 and Mythos 5 globally. The stated basis was national security.
Was OpenAI’s GPT-4o retirement the same kind of event?
No. OpenAI’s GPT-4o retirement was a provider decision tied to product and infrastructure planning, while the Anthropic shutdown followed a government directive. The effect for users was similar: access to a relied-on model ended.
Do customers own hosted AI models?
In most hosted AI services, customers receive permission to use a model through a product interface or API. They do not receive the model weights or the right to keep a specific hosted model online indefinitely.
Do open-weight models avoid this risk?
They reduce provider-deprecation risk if a team hosts them itself, because no outside provider can retire that exact deployed copy. They still depend on infrastructure, security maintenance, licensing terms and in-house ability to run them reliably.
What should AI teams do now?
Teams should inventory where models are hardcoded, test fallbacks before a failure, monitor provider deprecation pages and price changes, and design AI features so one unavailable model does not stop a core workflow.
Source: Thorsten Meyer AI
