Anthropic Locks Every Foreign User Out of Its Two Best Models

01The Friday-night order that locked every foreign user out of Anthropic's top two models

On Friday evening the US government ordered Anthropic to block access to Fable 5 and Mythos 5 for all foreign nationals, both inside and outside the United States, citing national security concerns. The directive reached past customers. It named Anthropic's own employees who are not US persons. To comply, the company cut off the two models entirely, for everyone.

Anthropic confirmed the suspension in a statement describing it as a US government directive to suspend access to Fable 5 and Mythos 5. The order arrived as an export control measure, the mechanism Washington normally reserves for weapons and advanced chips. Here it was pointed at API access to two language models.

On one side sits that national security rationale. On the other sits a global customer base and an internal staff that woke up Saturday without access to the products they build. The government drew no line between a foreign adversary and a foreign engineer on Anthropic's payroll. The blackout is total because the order left no narrower way to satisfy it.

What pushed a model from product launch to export-controlled is where the conflict sharpens. According to the Wall Street Journal, as reported by The Verge, the directive was triggered in part by cybersecurity research from Amazon, alongside conversations between Amazon CEO Andy Jassy and the White House. The Journal reported that Amazon's paper claims that, through a series of steps, the models could be misused. The full chain of those steps was not detailed in the report.

That sourcing puts a competitor's research at the origin of a government action against Anthropic's flagship models. The Verge attributes the account to the Journal, and Anthropic's own statement does not name Amazon or describe what the research found. The company is left enforcing an order whose stated basis it has not publicly detailed.

The practical effect is immediate. Developers building on Fable 5 and Mythos 5 lost their most capable Anthropic models overnight, with no carve-out for US-based foreign staff. A research paper from one company became the documented input to an export control order against another. The next signal to watch is whether the suspension is reviewed or made permanent, and whether other frontier labs face the same treatment under the same authority.

02"Subscription Economy for Cognition" Lands 1,500 Hacker News Points in a Day

A single-page document at opensourceaimustwin.com, signed by @TheAhmadOsman, climbed to 1,500 points and 463 comments on Hacker News. Its argument is blunt: if intelligence becomes "something people can only rent from a few closed institutions," the public loses not just software freedom but "operational freedom" — the ability to study, deploy, audit, and run models without asking permission. The post warns that frontier capability concentrated in a handful of labs risks turning core infrastructure into "a subscription economy for cognition."

Manifestos are cheap. The number behind this one is the signal: 1,500 points puts it among the day's most-voted items on a forum where most launches stall in double digits. The sentiment found an audience.

The same window gave that sentiment technical cover. MiniMax, an open lab, published two frontier methods directly to Hugging Face rather than gating them behind an API. MiniMax Sparse Attention drew 101 upvotes for a blockwise sparse design that targets the quadratic cost blocking million-token context. MaxProof, at 74 upvotes, describes a population-level test-time scaling framework for competition-level mathematical proof, with the generator, verifier, and refiner merged into a single released model.

Both papers attack problems — ultra-long context and rigorous proof — that closed labs treat as flagship differentiators. Releasing the methods publicly reframes the open camp's pitch. It is no longer only about freedom in principle. It is about whether a developer can read the design, reproduce it, and deploy it locally without a vendor's permission.

The timing sharpens the contrast. In the same period, the US government moved to cut off access to a closed frontier model on national-security grounds, a reminder that rented capability can be revoked by parties the user never contracted with. An open weight cannot be switched off by a shifting terms-of-service page.

The manifesto asks readers to "send a quiet note" to help build the case. The papers ask readers to download and run. One supplies the slogan; the other supplies the evidence that the slogan has working code behind it.

03Your iPhone Can Now Fill In Parts of a Photo You Never Shot

Open the editor in iOS 27 and the iPhone offers to do things a camera was never supposed to do. Reframe widens a cropped shot, painting in scenery that was never in front of the lens. Extend stretches the edges of an image past where the sensor stopped. Clean Up erases a stranger, a power line, a whole limb. In The Verge's hands-on test, the tools mostly work.

That is the unsettling part. Apple has put the first set of serious AI photo edits onto what The Verge calls the most popular camera in the world, and the reviewer's verdict came with a caveat: for better and worse. A photograph in iOS 27 can be reconstructed, expanded, and scrubbed before it ever leaves the phone. The results look convincing enough that a viewer cannot easily tell which pixels came from light and which came from a model guessing.

Apple kept the features tame next to rivals. Google's Pixel phones already swap faces between frames and generate objects from a text prompt. Reframe and Extend stay closer to what was plausibly in the scene, which is the point. The company that spent years marketing the iPhone as a tool for honest memories is now shipping software that quietly rewrites them, and most users will never toggle it off.

The same release supplied Apple a second, unfamiliar result: Siri works. For roughly fifteen years the assistant lived somewhere between mildly useful and unable to set a timer reliably. The new version handles real requests, according to The Verge, and the surprise alone made it the story. Apple's consumer AI has moved from punchline to something people might actually use on purpose.

For the person holding the phone, the practical change is narrow and concrete. Editing a photo no longer means cropping and adjusting light. It means deciding whether to keep what the camera saw or accept what the phone invented, with no label on the file to mark the difference. The tools ship to everyone who installs the update.