This isn't coincidence, it's a trend
June 12: Anthropic forced by US export control order to shut down Mythos 5 and Fable 5.
June 26: OpenAI required to phase-release GPT-5.6, with individual user approval.
Two weeks. Two of the world's most powerful AI companies, both hit by US government intervention. This isn't coincidence.
Anthropic: the harder hit
June 9, Anthropic released Fable 5 — a Mythos-tier flagship. Thought it was good news.
Three days later, a US export control order forced Anthropic to emergency-shut down Mythos 5 and Fable 5. All users blocked. Even Anthropic's own non-US-citizen employees were denied access.
June 17, Commerce Secretary Lutnick sent Anthropic a letter: before offering flagship model access to any foreign national, you must obtain a Commerce Department license. Violations face criminal and civil penalties.
This is the most substantive direct intervention into AI company operations by the US government to date. Not a suggestion. Not a voluntary framework. A legal threat.
Anthropic argued the reported "jailbreak" was narrow in scope, and GPT-5.5 could do the same thing. The government didn't listen.
OpenAI: better conditions, same direction
GPT-5.6's release process looks completely different from before.
No global ChatGPT rollout. No open developer API. Instead: "limited preview" — a small group of trusted partners gets API and Codex access. Partner names shared with the government.
The White House's National Cyber Director and Office of Science and Technology Policy directly intervened, requiring phased rollout. Altman told staff in an internal meeting this was the best path to broad release eventually, and employees must comply.
At least OpenAI could still release. Anthropic got shut down entirely.
All three tiers flagged high risk
The most striking safety label on GPT-5.6: Sol, Terra, Luna — all three models flagged "High Risk" in cybersecurity and bio/chemical domains. Including Luna, the cheapest and smallest.
Previously only flagship models got this label. Now even lightweight tiers are flagged, showing government concern about "frontier model weaponization" no longer distinguishes between capability tiers.
Why these two companies
It's not random. Anthropic and OpenAI are currently the world's strongest in code, cybersecurity, and agentic execution capabilities.
When model capabilities concentrate in these domains — writing code, finding vulnerabilities, autonomously executing complex tasks — it's no longer just a chatbot. It's technology that could affect real-world security.
Once viewed that way, release authority rarely stays with the company.
Timeline: from release to shutdown in 3 days
Anthropic's timeline is particularly worth examining:
- June 9: Fable 5 released, globally available
- June 12: Export control order issued, emergency shutdown within 72 hours
- June 17: Commerce Secretary letter, threatening criminal penalties
- Late June: Fable 5 possibly restored, Mythos 5 still banned
From release to shutdown: 3 days. That speed tells you two things: the government is already continuously monitoring frontier model releases, and they have a ready-to-deploy instant response mechanism.
Impact on Chinese users
The export control order's core provision: prohibit model access to "foreign nationals." This means Chinese users — whether enterprise or individual — may still be blocked even after Fable 5 resumes. Access likely requires routing through US entities, or waiting for open-source alternatives.
Short-term, DeepSeek V4 (MIT open-source) and Qwen3 remain the most reliable high-performance options for Chinese-language users.
Fable 5 might return this week
Good news: Axios reports the US government is about to lift Fable 5's export controls, with public access possibly restored this week.
But Mythos 5 remains banned. And "about to lift" doesn't mean "permanently safe" — today's loosening could tighten again overnight if another jailbreak incident occurs.
OpenAI's position
OpenAI stated clearly: they don't accept government pre-review as a permanent default mechanism. If the strongest tools are always delayed, users, developers, and cyber defenders get the best tools later.
Short-term compliance is to advance eventual full availability, while co-developing a reusable standardized release process with the government.
Nice words. But in execution, the transfer of model release authority is already happening.
Two threads advancing simultaneously: stronger technology → tighter government → more restricted companies. Anthropic's shutdown is the extreme case. OpenAI's limited release is the milder version. Same direction.
For regular users, the short-term impact is simple: "the newest, strongest models might not be available to you yet." For the industry, what changes is the entire release cadence — from "build and ship" to "build, review, then ship."
This won't be the last time. Next-generation Fable 6 and GPT-6 will face even more mature, institutionalized approval processes. The era of free frontier model releases has officially ended.