The block hit on a Tuesday. I was in a Slack room with about two hundred developers when someone’s phone went off, then someone else’s, and then we all just sat there. Nobody said anything for maybe thirty seconds.
First it was confusion. Then three companies posted the same “pausing operations pending review” statement within the same hour. Nobody had drafted that in advance. They just all had the same lawyers.
One team I knew well — a startup, maybe forty people — had bet their whole workflow on Mythos. Three months of work built around it. They had forty-eight hours to pull it. They did not make it.
Howard Lutnick does not usually send personal letters. So when one showed up at Anthropic’s chief compute officer’s desk two weeks after the block — that told me something was happening.
Amazon pushed back publicly. Their take: restricting access does not make anyone safer; it just sends people to less regulated options nobody’s monitoring. Fine. Amazon also wants to use the model. Weight their opinion accordingly.
Friday. The letter. “Appropriate safeguards in place.” The model can ship again, to a vetted list.
Nobody knows what that means in practice. The letter says it. The press release says it. What Anthropic has to actually do, report, prove — everyone’s guessing. “Appropriate safeguards” is the kind of phrase that means exactly what you need it to mean until suddenly it does not. I keep thinking about what happens the first time one of these companies gets audited. Or does not. There is a difference. We do not know which one is coming.
One thing nobody mentioned: Claude Fable 5. The smaller model that used to hold the “most powerful AI for consumers” title before all this started. Sources say it is coming. No timeline. “No clear timeline” has been the most honest sentence in AI policy for two years.
Here is what I keep circling back to: Anthropic went from subject of regulation to partner in it, in about fourteen days. That is fast. Whether it works for OpenAI, Google, whoever is next — I do not know.
GPT-5.6 Sol dropped the same day. Same story. Same oversight.
For enterprise teams: more paperwork, longer wait times, more people in the room asking what you are actually using this for. The total lockout scenario looks less likely than it did two weeks ago. Whether that is reassuring depends on how much you trust “appropriate safeguards” as a meaningful constraint — and honestly, that number changes depending on what week it is.
I did not expect the reversal. Did not expect it this fast. I was wrong about that.
The broader pattern here is worth sitting with: we are watching the early stages of AI governance becoming actual governance. Not frameworks, not principles, not voluntary commitments — actual negotiated agreements with real consequences and ongoing oversight. That is new. And whether it works is a question nobody can answer yet.
