The team you don't own

At 5:21 on a Friday afternoon, Anthropic got a letter. By the next day its most capable model, Fable 5, three days after launch, went offline. For everyone, worldwide.

A US government directive barred all foreign nationals from using it, under export controls. Anthropic can't verify citizenship in real time, so it shut the models off globally — for paying customers, even for its own foreign employees. Anyone mid-session got an error; their requests quietly dropped to a weaker model.

Fable 5 led every coding benchmark — twenty-two points ahead of GPT-5.5 on SWE-Bench Pro. The kind of gap you feel on agentic work. It was available for three days.

This hit me directly. I'm not American — the best tool on the market just got locked behind a passport. But that's not the part that keeps me up.

You build it, but you rent it

I've written before that I don't replace a team; I assemble one from scratch, out of agents, skills, memory. I stand by it. The team is real, and I work with it every day.

What this week reminded me is where it lives. Not on my machine. On rented compute, behind a model whose switch I don't hold. And as it turns out, the company I have a contract with isn't the only one holding it: a government I never signed anything with holds it too.

The agent I hand the work to is mine for exactly as long as some third party allows.

What you actually own

The answer isn't to retreat to my own hardware and pretend frontier models don't exist. A local, open-weights model buys sovereignty, and the price is the exact capability that made delegating worth it. This isn't loyalty. It's an engineering call.

The sane middle isn't ownership, it's substitutability. Building the system so it survives a model swap. Because what you actually own isn't the weights sitting in someone's datacenter; it's the context, the judgment, the structure of the work. The layer that's still there when the API throws an error underneath you.

The model is just access. The work is mine.