The US government is gutting Anthropic's R&D capacity
A Commerce export control pulled Claude Fable 5 and Mythos 5 on Friday. It's cutting roughly a third of Anthropic's own researchers off from the frontier models they build, while every competitor's keep theirs.
Anthropic released Claude Fable 5 last week. I used it for three days before the government switched it off.
I spent Tuesday through Thursday at a convening and Friday traveling without connectivity, so I worked at about a quarter of my usual pace. In that window I pointed Fable at a range of projects, including the microdata pipeline that PolicyEngine’s tax and benefit modeling runs on. That pipeline combines household surveys, administrative records, and calibration against published totals into a synthetic population for federal, state, and local analysis. We had spent months rebuilding it. Fable redesigned the architecture across the dependent packages, rebuilt the pipeline, and produced a dataset about five times more accurate than the one it replaces, measured against our held-out targets. That work reaches users this week. In the same three days it re-engineered the agentic flows behind several systems we ship over the coming weeks.
That is one model, used part-time, for three days. The effect that matters most from Friday’s order is not on its customers. It is on Anthropic’s own researchers.
What the order does
On Friday the Commerce Department directed Anthropic to suspend access to Fable 5 and the related Mythos 5 model for any foreign national, inside or outside the United States, including the company’s own employees. Commerce used the Export Administration Regulations’ “deemed export” rule, which treats releasing controlled technology to a foreign person in the country as an export to their home country. Anthropic could not screen its customers by nationality in real time, so it pulled both models for all users. Claude Opus 4.8 and the earlier models stayed up.
That global cutoff is the commercial cost. The research cost is narrower and more specific.
Who it bars
Anthropic knows each employee’s status, so the cut is precise: its foreign-national researchers lose access to the models they build, while their citizen and green-card colleagues keep using them. The deemed-export rule the order invoked exempts citizens, green-card holders, and “protected individuals”; the bite lands on employees on temporary visas. (One outlet read the order to cover green-card holders too; the rule it invoked says the opposite.)
How murky even that line is showed up within a day. Headlines reported that Andrej Karpathy, who joined Anthropic in May to lead its recursive self-improvement work, had been locked out of the models he was hired to build for lacking US citizenship. But the claim that he was barred traced to a single viral post sourcing his immigration status to an AI chatbot — no one verified it, and if he holds a green card the order never reached him at all. For days no one could say cleanly whether one of the field’s best-known researchers was in or out.
How many
Frontier AI runs on immigrant talent: about two-thirds of top-tier AI researchers in the United States earned their undergraduate degrees abroad, and international students are the core of the US AI PhD pipeline. No public figure exists for Anthropic, so I asked Claude to estimate it — five independent passes, each from a different angle: the AI-talent pipeline, Anthropic’s visa filings, the China–India green-card backlog, the company’s youth, and peer-company workforces. They ranged from 27% to 42% of its researchers, with a mean near 36% — on the order of a third. The passes agree on why it runs high: Anthropic is only four years old, so few foreign-origin researchers have had time to naturalize, and green-card backlogs for Chinese and Indian nationals — the field’s two largest talent sources — run a decade or more, leaving even long-tenured staff on temporary visas. It is an estimate, not a count: Anthropic publishes no breakdown, and its H-1B filings are flow, not stock.
The barred researchers also have an obvious exit. The control attaches to Anthropic’s models, not to the people, so a visa-holder cut off inside Anthropic can join OpenAI, Google, or any other lab and use a frontier model there on day one — every researcher at those labs, visa holders included, keeps access, because the order names only Anthropic’s models. So it does not just bench a third of Anthropic’s researchers; it hands every competitor a standing recruiting pitch aimed squarely at them, turning a model restriction into a talent-export pump pointed at one company.
Even if it’s temporary
None of this needs to be permanent to matter. Prediction markets expect a temporary pause — both Polymarket and Kalshi have access most likely returning within a few weeks. But the field doesn’t pause with it, and it’s speeding up: the length of tasks frontier models can do reliably has been doubling every seven months — lately closer to four, a doubling time that keeps shrinking. On a curve that steep and still accelerating, weeks on the sidelines — while rivals keep training, shipping, and hiring — shift leads and define careers. A temporary pause here is not a small one.
The incentive it sets
Applied as a standard, this penalizes shipping. Every frontier release would cut a lab’s visa-holding researchers off from the model they just shipped, the one they need to build the next. A lab that ships nothing keeps its researchers’ access; a lab that ships loses it.
The path back
A mechanism exists. The deemed-export rule routinely lets foreign nationals work on controlled technology under an individual license, usually backed by a technology control plan; semiconductor and aerospace firms employ non-citizens on restricted work this way. But BIS reviews each license under the rules for exporting to that employee’s country of citizenship. For nationals of close allies, approval is routine. For nationals of countries the United States restricts, including China, the review carries a presumption of denial. The route back is per-employee and country-by-country, measured in months, and it would return some of the workforce while leaving the rest in review.
On a consistent standard
The order’s authority is the Export Control Reform Act of 2018, applied through an “is-informed” letter to a single company — the first time it has reached an AI model — and Lutnick gave no basis for why Anthropic specifically. Legal analysts called it “not clear why Anthropic’s models were singled out … as compared to other U.S. large language model AI companies,” and “a remarkable contrast” with the administration’s hands-off stance on AI exports.
The directive names only Anthropic’s two models, and the jailbreak behind it is disputed. Anthropic says the government’s evidence was verbal, and demonstrated by a competitor, and that the same capability is available in OpenAI’s GPT-5.5, unrestricted. Security researcher Katie Moussouris described the technique as the gap between asking a model to “review code for security issues” and to “fix this code,” and said it “cannot meaningfully be fixed, and any attempt would only weaken the model for defense.”
The fixation on Anthropic predates the order. On a May 8 podcast — weeks before Anthropic overtook OpenAI in valuation — the administration’s AI czar, David Sacks, called it “the most powerful monopoly ever created in human history” and likened it to Standard Oil rebranded as “Safe Oil,” casting its safety record as cover for a monopoly. On the same show, Sacks said the cyber capability was not Anthropic’s alone — “OpenAI now has a model that’s just as cyber capable as Mythos,” with every major lab and Chinese models to follow within months — and his prescription then was industry-wide hardening, not a restriction on one company.
David Sacks’s “Safe Oil” monologue — All-In, May 8, 2026
“Unless something about their current trajectory changes, Anthropic will be the most powerful monopoly ever created in human history. […]
I just want you to think for a second about the case of John D. Rockefeller, who I think is known as probably the most successful, most ruthless monopolist in American history. But he wasn’t very good at PR. He was terrible at PR. Everyone sort of recognized how ruthless he is. We’ve seen movies like There Will Be Blood, which is basically about him.
In any event, imagine if John D. Rockefeller was way better at public relations, and instead of calling his company Standard Oil, he called it Safe Oil. … He called it Safe Oil because, as we know, kerosene is dangerous. Their first big product was kerosene. And kerosene can light your house or it can burn it down, and in the wrong hands it can torch a city or you can use it to make a bomb. So John D should have called for the creation of a new government agency to regulate the safety of his product, and they could have done rigorous testing, licensing, common-sense regulation. … And I think people would have gotten so wrapped up in this debate over what constituted Safe Oil or Safe Kerosene that they would have missed what was really going on, which is that Rockefeller was building the richest, most powerful monopoly of all. In fact, people might even have called Rockefeller an Effective Altruist, because of course he was so concerned about the safety of his product.”
The day before the export control, the same hosts opened their show attacking Fable for the opposite sin — too much caution. They called its guardrails “Orwellian” and its 30-day data retention “mandatory surveillance,” and Sacks called the safety posture “a very sophisticated regulatory capture campaign based on fear-mongering.” Later in the same episode, asked about Bernie Sanders’s plan to seize a government stake in the AI companies, Sacks said he “may be okay with Bernie’s idea” — but only for “a public benefit corporation that says it’s going to cause massive job loss, that trained for free on humanity’s knowledge but gatekeeps and refuses to give back,” and “maybe it should be 75%.” Sacks offered those as principled conditions, but they fit OpenAI as well as Anthropic: both are public benefit corporations, both trained on the same public data, both keep their weights closed, and both CEOs have predicted mass job loss. A test that singles out one of two near-identical companies isn’t a test; it’s a target with criteria attached. The next day the government restricted Anthropic, and only Anthropic, for being insufficiently safe.
No published standard tells another lab what to avoid, or tells Anthropic what to fix — only that, for now, the one company singled out must route its own researchers through a months-long license to use the model they built, or watch them rebuild it somewhere else.
The deeper irony is that standards-based, politically neutral governance is what Anthropic spent years asking for. It proposed uniform disclosure rules for every frontier lab in “The Case for Targeted Regulation”, and days before the order, Dario Amodei’s “Policy on the AI Exponential” argued that any government power to block a model “must be scoped to … four specific risks” with “protective measures against political favoritism or arbitrary decisions.” Defending the order, Sacks wrote that Anthropic “asked for government regulation of Mythos” — but its proposals were model-neutral, applied to all frontier developers, and it had gated Mythos itself rather than ask Washington to single it out. Anthropic asked to be governed by a standard applied to everyone; it got an ad-hoc decree applied to it alone.