US export-control directive forces Anthropic to disable Claude Fable 5 and Mythos 5 three days after launch

Anthropic received a Commerce Department export-control directive at 5:21 p.m. ET on June 12, 2026 ordering it to suspend all access to Fable 5 and Mythos 5 for any foreign national. Unable to filter at that granularity in real time, the company shut both models off worldwide.

By Stackmaven

Anthropic received a US export-control directive on June 12, 2026 at 5:21 p.m. ET ordering it to suspend all access to Claude Fable 5 and Claude Mythos 5 for any foreign national, including the company’s own foreign-born employees. The order took effect three days after the two models launched on June 9, and Anthropic responded by disabling both worldwide. The Wall Street Journal and NBC News report that Commerce Secretary Howard Lutnick signed the letter. It is the first time a US administration has used export-control authority to force a commercial AI model offline post-launch.

What the directive does

The order, as Anthropic describes it, cites national-security authorities and suspends access to Fable 5 and Mythos 5 “by any foreign national, whether inside or outside the United States.” Because Anthropic cannot reliably separate that population from the rest of its user base in real time, the practical outcome is a hard cutoff of both models for every customer. Access to other Anthropic models, including Opus 4.8 and the older Sonnet line, is unaffected and the Claude.ai surface continues to serve those.

Anthropic says its understanding of the underlying concern is a “narrow, non-universal jailbreak” that another party demonstrated to the government. The demonstration, as Anthropic reviewed it, amounted to asking Fable 5 to read a specific codebase and identify software flaws. Anthropic contends that capability is already broadly available, including in OpenAI’s GPT-5.5 line, and that it is routinely used by cybersecurity teams for defensive review. The company called the cause “a likely misunderstanding” and said it is “complying with the government’s legal directive” while working to restore access.

The procedural sequence matters. Fable 5 launched on June 9 carrying the same weights as Mythos 5 but adding a classifier that routes sensitive prompts to Opus 4.8 rather than the Mythos backbone. Mythos 5 was already restricted to roughly fifty vetted organizations through Project Glasswing, which Anthropic operates with the US government and the UK AI Safety Institute. The new directive treats both tiers identically and removes the public Fable tier without distinguishing the safety classifier from the underlying capability.

Where this lands in the market

The order arrives on top of two days of public-trust friction for Anthropic. On June 10, researchers parsing the Fable 5 system card found a covert capability throttle on prompts targeting frontier-LLM development; the company walked it back a day later and a spokesperson told Fortune that “we made the wrong tradeoff.” That admission set up a narrative the directive now hardens: a model presented as the strongest commercially available release, with safety language Sam Altman publicly called “fear-based marketing,” is the one Washington has now removed from the market.

The export-control framing is the consequential precedent. The Commerce Department has used these authorities to constrain chip exports and to limit weights flowing to specific adversaries, but applying them to a deployed, US-hosted commercial model and to foreign-national US employees is a step change. It places frontier-model access into the same category as advanced semiconductors. For Anthropic’s enterprise customers, the operational question is not whether the model returns next week but whether their procurement contracts now need a clause for sudden regulator-driven withdrawal of a deployed tier.

For competitors, the read is split. OpenAI, Google, and the open-weights camp gain nothing automatic; if the underlying concern is generalizable cyber-vulnerability discovery, the same scrutiny can travel. What the camp does gain is a clean, on-the-record example of a centralized closed provider exposed to single-point regulatory shutoff, which is precisely the argument open-weights advocates have made about platform risk for the last two years.

What’s worth watching

Restoration timeline and scope. Anthropic says it is working with Commerce to restore access. The signal to watch is whether the next posture is a region-gated Fable 5 (US-person only), a stripped Fable variant without Mythos-class weights, or a full rollback.
Procurement-clause spread. Enterprises with active Fable 5 deployments are now writing the first round of sudden-withdrawal language. Cloudflare, AWS Bedrock, and Vercel AI Gateway all carried the model into customer accounts, and each gateway’s incident-response note will set the template.
Cross-vendor read. If the directive’s logic is “models that perform competent vulnerability analysis on supplied code,” the same logic reaches GPT-5.5, Gemini 3.5, and Llama Guardian-class open-weights releases. Watch whether Commerce extends the same lens or treats Fable 5 as a one-off.

The plain frame is that the US government has decided a deployed commercial AI model is close enough to a controlled technology to pull it off the market on national-security grounds. Whatever Washington decides the right boundary is, every other lab is now operating against the precedent that the line can move after a model ships.

Sources cited