← All writing
AI Jul 1, 2026 6 min read

Anthropic's best model came back. Government pre-clearance came with it.

Christopher Dorsey

Christopher Dorsey

AI & MadTech Advisor · Enterprise Sales Leader

TL;DR

On June 30 the US Commerce Department lifted the export controls that had taken Anthropic's Fable 5 and Mythos 5 dark for nearly three weeks, and Fable 5 began rolling back out worldwide on July 1 under a capped quota through July 7. But in the same fortnight, OpenAI previewed GPT-5.6 Sol only to government-vetted "trusted partners" at Washington's request. Government pre-clearance of frontier models is becoming standard, built on the decades-old "deemed export" rule that took Fable 5 down globally because a control aimed at foreign nationals became an outage for everyone. Multi-cloud doesn't help — the restriction attaches to the model, so AWS, Google Cloud, and Foundry all went dark together. If you sell AI on a frontier model, your buyer inherited its regulatory exposure: build a real fallback, keep workflows portable, and raise it before the security review does. If you buy, add regulatory interruption to your evaluation and ask where the model is cleared and how fast it was restored last time.

On June 30 the US Commerce Department lifted the export controls it had placed on Anthropic's Fable 5 and Mythos 5, and Fable 5 started rolling back out worldwide on July 1, at first under a capped usage quota through July 7. The two models had been dark for nearly three weeks. On June 12 the government barred any foreign national from accessing them, and Anthropic, with no way to verify nationality at the login screen, shut them off for everyone rather than risk breaking the rule. When it happened I wrote that a frontier model disappearing in 96 hours by government orderturns model continuity into a procurement category. The model is back. The reason it went away hasn't changed, and in the same stretch it showed up at a second company.

What actually got approved

When OpenAI previewed GPT-5.6 Sollast week, it said it had shown the government both its rollout plan and the model's capabilities before launch and, at the government's request, was releasing it first only to a small group of trusted partners whose names it had shared with Washington. Anthropic got Fable 5 back after Commerce and its Center for AI Standards and Innovation reviewed a retrained safety classifier. Two of the three frontier labs cleared a top-end release with the US government inside the same two weeks. A government sign-off before the best models reach the market is no longer a one-off tied to a single jailbreak. It is turning into the way frontier models ship.

The mechanism analysts point to is the “deemed export” rule, the decades-old idea that giving a foreign national access to controlled technology counts as exporting it. Applied to a model, that is what took Fable 5 down everywhere at once: a control aimed at foreign nationals became an outage for every user, because Anthropic couldn't check nationality in real time. As one analyst put it, the reversal isn't the story; the instrument underneath it is. Washington now has a working lever on which model reaches which users, and it has pulled it twice in a month.

Why multi-cloud doesn't save you

The reflex for an enterprise worried about a single vendor is to spread the risk across providers. It doesn't help here. Fable 5 runs on AWS, Google Cloud, and Microsoft Foundry, and all three went dark together, because the restriction attached to the model, not the cloud it sat on. When the constraint is federal policy aimed at the model maker, buying the same model through three vendors gives you three copies of the same exposure. The thing worth diversifying is the model itself, not the invoice.

If you sell

For anyone selling AI built on a frontier model, this belongs in the security review, and you should raise it before the buyer does. Your customer inherited your model provider's regulatory exposure the day they signed. The answer that holds up isn't “it hasn't happened to us” — it happened to the most capable model on the market, and a capped-quota relaunch is not the same as full availability. Build a real fallback path to a second model. Keep your prompts and workflows portable, so a switch is a config change instead of a rebuild. Be able to say, in plain terms, what a customer's Tuesday looks like if your primary model gets gated for a week. That's a feature now, not a footnote.

If you buy

Add regulatory interruption to the checklist next to security and price. Ask where the model is cleared, who counts as a permitted user, how fast the vendor restored access last time, and whether you'd make the cut for “trusted partner” early access or wait at the back of the line. The Fable 5 outage ran almost three weeks, and the relaunch is throttled for a week after that. Multi-cloud won't insulate you from a rule written against the model maker, so treat continuity as its own line of diligence rather than a box the cloud contract already checks.

In June the takeaway was defensive: keep a fallback, because a model can vanish. Two weeks later the picture is larger than one company's bad fortnight. The government has shown it will clear frontier releases before they ship and pull them after, and the labs are building that review into how they launch. The model you standardize on is now governed in part by a policy decision you don't control and can't see coming. Restored access isn't restored certainty. Price the detour into the contract, because the road runs through Washington now.

Share this post

About the author

Christopher Dorsey

Christopher Dorsey

Enterprise Sales Leader · AI Go-To-Market · Startup Advisor · Denver, CO

Fifteen years selling technology to Fortune 500 brands across AI, advertising, and data infrastructure — most recently at Zeta Global, Oracle, and Fastly. Currently advising founders and sales leaders on AI go-to-market and Generative Engine Optimization.

Questions, pushback, or just want to compare notes?