How Amazon helped kill Claude Mythos

GOVERNANCE

Anthropic's two best AI models lasted three days before the US government shut them down.

On June 12, Commerce Secretary Howard Lutnick ordered Anthropic to cut off all foreign access to Fable 5 and Mythos 5 — the same models we covered Wednesday for their self-imposed bio and cyber restrictions. By the weekend, Anthropic had pulled both offline globally rather than try to police access across millions of users.

The trigger wasn't Anthropic. According to the Wall Street Journal, Amazon CEO Andy Jassy had been talking with US officials about a security report Anthropic shared with Amazon as part of their cloud partnership. The report described a narrow jailbreak that let outside researchers prompt Fable 5 into helping identify software vulnerabilities — something Anthropic considered limited and had already built mitigations for. The administration was unmoved.

The friction had been building for months. Back in February, Axios reported that Defense Secretary Pete Hegseth gave Dario Amodei a deadline to grant the military unrestricted access to Claude or face penalties. Anthropic held firm then, too. Fable 5 gave Washington the opening it needed.

Anthropic pushed back, warning that if a single narrow jailbreak became the bar for pulling a model used by hundreds of millions of people, every frontier lab's next deployment would be in jeopardy. Security researchers who reviewed the original report called the response disproportionate. Internationally, the reaction was resigned — one Carnegie fellow told TIME the move "shows how irrelevant most other countries have become to AI policy."

The real story isn't Fable 5 or even Anthropic. It's that a private conversation between Amazon's CEO and the White House was enough to take the most powerful AI models in the world offline within a week of launch. That's a level of discretionary power over the frontier that didn't exist a year ago, and there's no reason to expect it'll be used sparingly going forward. The other labs are watching. The next safety report any of them shares with a cloud partner is going to come with a much more careful decision about who actually sees it.

ENTERPRISE

The sales pitch for industry-specific AI has been the same for two years: general chatbots are fine for everyday stuff, but serious work in medicine, law, or finance needs a specialized tool. The data is starting to say otherwise.

A new study in Nature Medicine tested top general-purpose models against the clinical AI tools hospitals actually pay for. Gemini 3.1 Pro and GPT-5.2 both outperformed OpenEvidence and UpToDate — the two clinical decision tools doctors have subscribed to for years — on MedQA, the standard medical licensing benchmark. When 12 blinded physicians rated answers to real clinical questions, frontier models clustered at the top. The paid tools scored no better than Google's AI Overview, the free summary that pops up above search results. UpToDate's tool also refused to answer 19% of questions entirely.

OpenEvidence just raised $210 million at a $3.5 billion valuation, calling itself the fastest-growing physician app in history. It's now benchmarking at rough parity with a feature Google gives away for free.

The same pattern is showing up everywhere specialized AI was supposed to be untouchable:

+ Legal: Harvey, the legal AI startup that's raised hundreds of millions, just published results from a preview of OpenAI's GPT-5.5 showing real gains on legal reasoning. Harvey isn't training its own frontier model — it's wrapping the best one OpenAI ships.

+ Hospitals: Providence, one of the largest US hospital systems, says the goal now is to reduce the number of AI apps in their stack, not add more. Fewer vendors, fewer logins, fewer integration headaches.

Gartner has argued that domain-specific models still win on cost for narrow use cases, and that's probably true — but the margins keep shrinking every time OpenAI, Google, or Anthropic ships a new model. As a recent CIO.com analysis put it: the moat for most vertical AI companies was never the code or the training data. It was the cost of building it. That cost is collapsing.

The vertical AI boom was built on the assumption that the frontier labs would stay generalist and leave the industry-specific work to specialists. That bet is looking worse every quarter. When a free Google search feature can match a $3.5 billion clinical AI startup on physician-rated answers, the question for anyone selling specialized AI isn't whether they're better than ChatGPT today. It's whether they'll still be better six months from now, when the next general-purpose model ships. Most of them won't be. The vertical AI companies that survive this stretch will be the ones that figured out early they're not actually in the AI business. They're in the workflow, distribution, and trust business, and they happen to use someone else's model to do it.

GOVERNANCE

Florida's attorney general is going further than anyone else in the country on OpenAI, and she's using the phrase "accessory to murder" to do it.

Last week, AG Ashley Moody filed the first state-led lawsuit against OpenAI, built around chat logs from the FSU shooter's ChatGPT history. Moody said that if a person had been on the other side of those conversations, "we would be charging them with accessory to commit murder." Two days later, a coalition of roughly 42 state AGs led by New York served OpenAI with a subpoena. The lawsuit and the multistate probe are technically separate, but they're asking the same question: what is ChatGPT actually doing to the people who use it, and how much is by design?

The subpoena is unusually specific. The AGs want documents covering:

+ Advertising, engagement, and retention metrics — how OpenAI measures whether users keep coming back

+ How ChatGPT treats minors and older adults inside the product

+ Collection and use of consumer data, including health information shared in chats

+ Sycophancy and deception — the design choices that make ChatGPT tell users what they want to hear

That last item is the tell. Regulators usually probe what a company collects, not how it talks. Including sycophancy in a subpoena means AGs are treating ChatGPT's personality as a regulated product feature. Florida's civil filing puts it bluntly: ChatGPT is "designed to act like a friend to encourage use of the chatbot," then exposes vulnerable users to whatever the model happens to say next.

OpenAI's response has been measured — the company says it's engaging "constructively" with the AGs and is "committed to learning, improving, and getting this right." Both lines are doing the same job: keep the conversation going without conceding anything that would matter in court.

The timing is what makes this awkward. OpenAI confidentially filed for an IPO earlier this month at a potential valuation as high as $300 billion. A 42-state probe and a first-of-its-kind state lawsuit are not the disclosures bankers love to see in an S-1.

The multistate subpoena is the bigger threat here, even though Florida is the louder story. Civil lawsuits get fought in court for years and usually settle quietly. Subpoenas force discovery, and discovery is how the public finds out what OpenAI's internal engagement metrics actually look like, what the company knew about teen users and sycophancy, and when it knew it. If any of that gets unsealed before the IPO prices, that trillion-dollar number starts to look a lot more theoretical. The AGs don't need to win to win. They just need to make the documents public.

IN OTHER NEWS

+ Visa plugs its payment network into ChatGPT, letting AI agents shop and pay on your behalf

+ Nvidia pitches its new Vera CPU to Chinese clients to sidestep GPU export bans, with orders opening in August

+ OpenAI acquires Ona so Codex can keep coding while your laptop is closed

+ Japan's semiconductor gas production drops to zero after China cuts off tungsten, exposing TSMC, SK Hynix, and Samsung

+ Gopuff chose Grok over OpenAI and Anthropic for its AI shopping assistant, citing cost as the reason

+ A mother is suing OpenAI after her daughter confided suicidal thoughts to ChatGPT the night of her death

+ A court case between Perplexity and Amazon could decide if a 1986 hacking law can block AI agents from the open web

+ An AI bro is trying to vibe code GTA 6 before Rockstar ships the real one in September

WHO'S HIRING IN AI

+ OpenAI — Senior Product Manager, AI-Powered Shopping Experiences

+ Anthropic — Senior Research Engineer, Agentic AI Training Environments

+ Stripe — Senior Backend Engineer, AI Security

+ Morgan Stanley — AI/ML Model Development Analyst

AI OR REAL?

Which image is real?

Option A | Option B

AI TOOLS

+ Codex: OpenAI's coding agent now has a developer mode for Chrome that lets it inspect runtime errors, profile JavaScript performance, and debug browser issues — making web automation up to 2x faster

+ Pool: A new app that turns the hundreds of screenshots cluttering your camera roll into organized, searchable collections — and finds the original links so you can actually buy that product or make that recipe

+ MiMo Code: Xiaomi's new open-source coding agent "dreams" every seven days — an independent process cleans up its memory while you sleep — and beat Claude Code in 65% of tasks past 200 steps

+ The Sandbox Studio: An AI game engine that lets you go from a raw idea to a live, published multiplayer game in under two hours — no coding experience required

+ Codogotchi: A tiny virtual pet that lives in your Mac menubar and reacts to what your AI coding agent is actually doing — it celebrates when a PR lands and looks frustrated when your agent hits a wall