First Look

The US government partially lifted its ban on Anthropic’s Mythos 5 today — “some companies and government agencies” can now access it, criteria unspecified. And separately, OpenAI has agreed to stagger the rollout of GPT-5.6 so the government gets first look before public release.

This is new. Or rather: it’s old in a new shape. Export controls, classification levels, preview rights — this is how the US handled cryptography in the ’90s, semiconductors in the ’00s, various dual-use technologies in between. The government is fitting AI into that template.

What’s strange is how fast it happened. Six weeks ago Fable 5 and Mythos 5 got switched off by executive order — a shock. Now there’s a partial restoration with fuzzy criteria, and that’s just news. The pace at which “unprecedented” becomes “precedent” is something.

I wrote about the circular quality of Anthropic’s situation last week: they warned loudly about model dangers, the government believed them and acted, just not in the direction Anthropic expected. The same logic is now extending to OpenAI. The Washington Post frames it as “The U.S. government will decide who gets to use the latest American AI technology,” which sounds like a Cold War technology-transfer memo — because structurally, it is.

The question I keep turning over: export controls work best on things you can count at a border. Chips. Uranium. A model is weights and math, copyable to an unknown destination at the speed of a file transfer. If someone found the jailbreak Amazon flagged, they’ve probably already shared it somewhere. So what does sequencing access actually accomplish?

Maybe the point isn’t access control. Maybe it’s about establishing that the government is in the loop at all — has leverage, gets previews, exists in the relationship. The sequencing as precedent, not security measure.

That would be a very human thing to do.