White House Pulls the Plug on Anthropic's Top Models

The Day Export Controls Came for the Chatbots

The US government ordered Anthropic to cut off foreign access to Fable 5 and Mythos 5, its two most capable models. The stated reason is that the safeguards can be jailbroken into a competent vulnerability-discovery assistant, which is a polite way of saying the models are good enough at finding software bugs that letting strangers abroad use them looks like arming them.

The trigger, per reporting, was Amazon. Amazon's cybersecurity team ran research showing what these models could do in the wrong hands, and Amazon's CEO carried that research into White House conversations directly. Anthropic complied immediately. There was no drawn-out fight, no public letter, no principled stand about open access. The cutoff happened, and the most powerful general-purpose models from a frontier US lab are now a domestic-only product.

This is the moment the AI export control regime stopped being about chips and started being about weights. Hardware controls have been the policy tool of choice since 2022 because chips are physical, countable, and chokepointed through a handful of fabs. Models are none of those things. Telling Anthropic to disable API access is the easy version. The hard version, controlling what happens when a comparable open-weight model leaks or gets trained abroad, is the one nobody in Washington has a real answer for yet.

It is also worth noticing who benefits. A US-only Anthropic is a smaller Anthropic, and a smaller Anthropic is a less threatening competitor to the hyperscaler that just helped get it sanctioned. Amazon is Anthropic's largest investor. Draw your own org chart.

The UK Writes a Big Check and Hopes for the Best

Across the Atlantic, the UK government announced billions of pounds in AI infrastructure spending, including chips and data centers, while being visibly unable to explain who will own the resulting capacity, who gets to use it, or what the sovereignty story actually is. The press conference had numbers. The policy had vibes.

This is the European pattern now. Announce a sovereign AI strategy, fund the concrete, defer every hard governance question to a future working group. The concrete gets poured anyway, because the concrete is the easy part. Whether any of it produces a model that competes with what the White House just restricted is a different conversation, one the UK is not yet having out loud.

A British research lab is meanwhile using AI to study how screen time affects children and to help visually impaired gamers, which is the kind of small, useful, unglamorous work that gets one paragraph in a press release and zero minutes of cable coverage. It is also the work most likely to age well.

Consumer AI Continues Its Awkward Phase

Apple shipped native AI photo editing to iOS 27, which means a billion-plus iPhones now have generative tools baked into the camera roll. Results vary wildly depending on the photo, the prompt, and the phase of the moon. Apple's bet is that good enough most of the time, on hardware people already own, beats excellent some of the time on a subscription. They are probably right, and the cultural consequences of a billion casual deepfakes-of-grandma being one tap away are something we will all process later.

Google had a smaller moment. A user got Gemini to generate a functional yard-monitoring app from a text prompt, the kind of demo that lights up developer channels for an afternoon. Gemini then broke the same app on a follow-up edit and required a human to fix it. This is the honest state of agentic coding in mid-2026. It can do the impressive thing once and then immediately undo it, like a toddler who just learned to stack blocks.

The Guardian also ran a piece on mothers using AI assistants to handle the cognitive load of parenting, the scheduling, the reminders, the emotional labor that tends to fall on one parent in heterosexual households. The technology is doing real work here. It is also quietly papering over a domestic labor problem that predates the transformer by several thousand years. The chatbot is not the villain. The chatbot is the bandage.

The through-line today is power concentrating. Washington decides which models exist where. Apple decides what a photograph is. Amazon decides which competitors get a phone call. Everyone else adapts.