TL;DR
Anthropic disabled Fable 5 and Mythos 5 worldwide after a June 12 U.S. export-control directive covered foreign nationals, including some of its own staff, according to the company. OpenAI’s February retirement of GPT-4o showed a separate provider-led path to the same result: users and developers can lose access to hosted models they do not control.
Anthropic’s June 12 shutdown of Fable 5 and Mythos 5 after a U.S. export-control directive, paired with OpenAI’s February retirement of GPT-4o from ChatGPT, has made model access a live dependency risk for businesses and users that build workflows on hosted AI systems.
Anthropic said it received a letter at 5:21 p.m. Eastern on June 12 citing national security concerns and applying to foreign nationals inside and outside the United States, including foreign-national Anthropic employees. Because the company said it could not reliably screen access that way in real time, it disabled both models globally, Business Insider reported.
Administration figures described the order as a response to a possible jailbreak of Fable 5 that could expose Mythos-class cyber capabilities. Anthropic has disputed that account, saying the concern was narrow and did not show capabilities beyond other public models, according to Tom’s Hardware and The Verge.
OpenAI’s move was different: it was a product retirement. OpenAI announced GPT-4o, GPT-4.1, GPT-4.1 mini and o4-mini would leave ChatGPT on February 13, and reports citing the company said GPT-4o accounted for about 0.1% of user selection at the time. OpenAI’s API deprecation page separately lists older model snapshots and preview systems scheduled for shutdown on fixed dates.
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.
AI Builders Face Access Risk
The two cases matter because they expose the same weak point from opposite directions. A government order can block a model for security or export-control reasons; a provider can retire, reprice, rate-limit or geofence a model as part of a product plan. In both cases, customers usually have access rights, not ownership of model weights or hosting.
For businesses, the risk is operational. A model string hardcoded into production can affect customer support, coding tools, data analysis, internal assistants and security workflows when it disappears or changes behavior. For individual users, the risk is personal as well, because some relied on GPT-4o’s conversational style for creative work, accessibility support or companionship, according to reports on user reactions after its February retirement.
LOCAL LLM DEPLOYMENT: Training, Fine-Tuning, & Offline Inference: The Complete Developer’s Guide to Building, Training, and Running Private Open-Source AI Offline (with full source code)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Two Shutdown Paths Converge
AI export policy had previously focused mainly on chips and compute. The Anthropic directive showed that deployed models can also become a control point when the provider is a U.S. company serving access through cloud APIs. The order’s reach was broader than a country block because it covered nationality, not only location.
OpenAI’s GPT-4o case had a longer path. The model was released in May 2024, removed after GPT-5 launched in August 2025, brought back after user backlash, then retired again in February 2026 with newer models offered as replacements. Reports from The Guardian and Business Insider described anger and grief among some users, while OpenAI said newer models were meant to carry forward preferred traits with stronger safety boundaries.
“We believe this is a misunderstanding.”
— Anthropic
Local AI with Ollama: Run, Customize, and Deploy Private Language Models on Your Own Hardware (Developer guides)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Security Evidence Remains Disputed
Several facts remain disputed. The government has not publicly released the full technical evidence behind the directive, and Anthropic says the letter did not give specific details. Reports differ on who first raised the alleged jailbreak concern and whether any China-linked actor had meaningful access to Mythos-class capability; Anthropic has denied that its safeguards exposed novel abilities.
It is also unclear how long the controls will remain, whether other U.S. model providers will face similar restrictions, and how regulators would enforce nationality-based access rules at scale. On the OpenAI side, the long-term question is whether providers will adopt longer notice periods or stronger legacy support for models that users build into products.
Epson RapidReceipt RR-600W Wireless Desktop Color Two-Sided Receipt and Document Scanner with ScanSmart AI PRO Receipt Management and PDF Software for PC and Mac, Touchscreen and 100 Page ADF
SCANSMART AI PRO TECHNOLOGY — Get your documents AI-ready1 with the help of the Epson RapidReceipt Receipt and…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Washington Talks Set Model Access
Anthropic is expected to continue talks with U.S. officials on a security review or patch path that could restore access. Customers using Fable 5 or Mythos 5 will need temporary model substitutions unless the controls are lifted.
For developers, the near-term action is to audit model dependencies, test fallback routes across providers, and decide where open-weight or self-hosted models are needed for continuity. Providers and regulators are likely to face pressure for clearer notice rules, appeal paths and technical standards for future model restrictions.
Self-Hosted AI Assistant for Beginners: Build a Private Open-Source Workflow with OpenClaw
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Did users own Fable 5, Mythos 5 or GPT-4o?
No. Customers accessed hosted services or APIs. They did not control the model weights, hosting layer or provider access rules.
Why did Anthropic disable the models for everyone?
Anthropic said the U.S. directive covered foreign nationals inside and outside the United States, including some employees, making selective compliance impractical in real time.
Was OpenAI’s GPT-4o retirement the same kind of event?
No. Anthropic’s shutdown followed a government order. OpenAI’s GPT-4o retirement was a provider decision tied to product support and newer model offerings.
Can developers prevent this risk?
They can reduce exposure, but hosted-model access cannot be guaranteed by users alone. Practical steps include fallback models, provider-neutral routing, regular migration tests and self-hosted models for workloads that need continuity.
Source: Thorsten Meyer AI
