How Powerful Is AI Now? Ask Washington

Earlier this month, the payments company Stripe took a fifty-million-line codebase, the kind of sprawling, decades-deep software that a team of engineers would normally spend two months rewriting by hand, and migrated the entire thing in a single day. The tool that did it was an AI model called Claude Fable 5. It had been available to the public for three days.

Three days is also how long it lasted. On Friday, June 12, the United States government ordered the model switched off for every person on the planet who is not American. To understand why a piece of software gets recalled like a contaminated product, and why that should interest Israel in particular, it helps to start with how capable these systems have quietly become.

Because the headline here is not a grievance. It is a snapshot of where artificial intelligence now sits: powerful enough that governments have started treating it as a strategic asset, and distributed unevenly in a way that matters most to the countries that lean on it the hardest. No country leans harder, per person, than Israel.

Consider what these models can now do. Anthropic, the company behind Claude, disclosed this month that more than 80% of the code it merges into its own systems is now written by Claude, up from low single digits two years ago. On the company’s hardest internal engineering problems, the model’s success rate climbed from 26% to 76% in six months. Last year, general-purpose AI models reached gold-medal standard at the International Mathematical Olympiad and earned a perfect score at the world finals of the premier university programming contest, outscoring every human team in the room. An independent lab that measures how long a task an AI can finish on its own has found that this duration is now doubling roughly every three months.

Fable 5, the model at the center of all this, was built to stay coherent across millions of words of context, which is what let it digest Stripe’s codebase whole. These are not chatbots that autocomplete sentences. They are systems that do multi-hour, expert-level work.

That power is exactly why the US government treated a flaw in Fable 5 as a national-security matter rather than a product bug. Anthropic itself said the high-end models behind Fable excel at discovering and exploiting software vulnerabilities, that they can make cyberattacks substantially easier and cheaper to commit, and that they had reached a threshold where they present significant risks. For that reason the company quietly routes sensitive cyber, biology and chemistry questions to a weaker, older model.

The problem, according to reporting, was that a trusted partner (Amazon, by several accounts) showed administration officials a way around that safeguard, coaxing the model into producing restricted offensive-cyber information it was built to withhold. The White House asked Anthropic to fix the hole or take the model down. When that did not happen on the government’s timeline, the Commerce Department reached for the bluntest fast-acting tool it had: an export-control directive barring access by any foreign national. Anthropic cannot verify the passport of every user in real time, so it switched the model off for everyone.

It is worth being fair about this. Anthropic disagrees that a narrow jailbreak justified pulling a model used by hundreds of millions of people, and notes the same information is available from rival systems. But from the government’s chair, the logic is coherent: a tool its own maker had certified as capable of cheapening cyberattacks turned out to have a safety layer that could be defeated, and the state moved to contain it. White House AI adviser David Sacks called the issue serious but easily resolved. This reads less like a crackdown than like a country discovering, in real time, that it now has to govern software the way it governs other powerful dual-use technology.

Which brings the story back to Israel. By Anthropic’s own Economic Index, Israelis use Claude about seven times more, relative to population, than the world average, the highest rate of any country on earth. That is not a quirk of national enthusiasm. Israel runs more than 2,300 active AI companies, and AI now draws close to half of all the venture funding flowing into Israeli startups. Even through a war, Israeli startups raised about $8.6 billion in the first half of this year, up roughly 45% on the year before, with the money concentrating in AI and cyber.

A great many of those companies build their products on top of frontier models made in the United States. That is the quiet stake in an episode like this one. When the most capable models are available to Americans but not, even for a few days, to foreign nationals, the builders most exposed are precisely the ones operating closest to the frontier. Israel, the heaviest per-capita user in the world, sits in exactly that position. If access to the best tools should ever become a question of nationality, the country that got the most out of those tools has the most to think about.

None of this is cause for alarm, and the episode cuts more than one way. The restriction was framed as temporary and fixable; the older Claude models kept running throughout; and open-weight models from Meta, DeepSeek and Mistral now trail the closed frontier by months rather than years, with Israel building its own sovereign Hebrew models at the Dicta project. On the hardware side, Israel’s standing with Washington has actually been drifting toward preferential, ally-level access. The frontier was nudged out of reach for seventy-two hours, not walled off.

Still, the lesson lands. Frontier AI has become capable enough, and valuable enough, that the people who build it and the governments that host it now treat access to it as something granted rather than assumed. The countries that do well in that world will be the ones that both use these tools intensively and invest in the capacity to build them. Israel already leads the world at the first. The week the most powerful AI on earth blinked off for everyone outside America was a useful nudge to get just as serious about the second.