3 minute read
The Quiet Part Nobody Voted On
Access to the best AI now runs through an approval list. That’s not a hot take, it’s basically how it works now. The U.S. government cleared Anthropic to release its top-tier Mythos 5 model, but only to a curated list of roughly 100 companies and federal agencies. Everybody else gets to stand outside the velvet rope and wait.
This is the first time the feds actually pulled a model offline
For years, AI releases worked on the honor system. Companies shipped what they wanted, when they wanted, and the only thing standing between you and a shiny new model was a credit card. That era just ended. The government slapped export controls on Mythos earlier in June over national security concerns, the model went dark worldwide, and the company spent two weeks in daily negotiations to get even part of it back. And to be fair, the worry wasn’t imaginary: researchers reportedly found a way to bypass the guardrails on Fable 5, the weaker model that briefly went public, and officials feared that crack could unlock the far more powerful cybersecurity capabilities sitting underneath in Mythos, which bad actors could then turn around to breach government and critical systems. Voluntary restraint is out. Permission slips are in.
Same day, OpenAI got the exact same treatment
The Mythos news didn’t land alone. The very same Friday, OpenAI dropped its new GPT-5.6 lineup and quietly admitted it was limiting access to a “small group of trusted partners,” also at the government’s request. So this isn’t Anthropic getting singled out for being naughty. It’s a pattern. Two of the biggest labs on the planet, both releasing frontier models on a government-approved guest list. When it happens once it’s a story. When it happens twice in one afternoon, it’s a regime.
“Trusted partner” is where everyone wants to be
Notice the language: trusted partners. That phrase didn’t exist as an access level a month ago, and now it’s the whole ballgame. There’s a list, the Commerce Secretary controls who’s on it, and he made a point of saying he can rewrite that list whenever he feels like it. So your access to the smartest tools on Earth now depends on whether your name made the cut on a letter you’ll never see.
So what do you actually do about it
Nothing dramatic. But pay attention to who’s on the list and who isn’t, because that list is about to matter more than any benchmark score. Which model you can actually use now depends as much on approval as it does on what the thing can do.
Sources: CNBC, Semafor, Reuters, The Hill