Trump Hosts Anthropic's CEO, Open Models Hit 56% at Vercel

Earlier this year the Pentagon labelled Anthropic a supply chain risk. This week Donald Trump privately hosted Anthropic CEO Dario Amodei for dinner. In quieter news, open models now carry 56% of Vercel's AI tokens. That second story is the one that shows up on your invoice.

One of these is politics. The other is procurement. Guess which one I think matters more to your business.

Trump hosted Anthropic CEO Dario Amodei after months of public fighting

Quick recap, because the background here is genuinely strange. The Pentagon asked to use Anthropic's Claude models for things like mass surveillance and autonomous weapons. Anthropic said no. The Pentagon then declared Anthropic a supply chain risk.

Trump called Anthropic's leadership left-wing nut jobs. Then a comedy sketch mocked Amodei's public warnings about AI. And now the two of them have had dinner.

Nobody has confirmed what was discussed. I have a hard time picturing a breakthrough.

One side says slow down. The other says build more of it, faster. Those positions don't meet in the middle over dessert.

What makes this less silly than it sounds is that the two sides reportedly agreed on almost everything when they met before. The sticking points were narrow. Mainly the weapons part. Anthropic's position is basically that today's models are not reliable enough to be pointed at anything that fires.

I happen to agree, and not for noble reasons. I've watched good models confidently do the wrong thing in a spreadsheet. The idea of the same behaviour with a targeting system is not a debate I need to have.

Meanwhile Australia has asked both Amodei and OpenAI's Sam Altman to appear at a Senate hearing. The request is not compulsory. So it is technically a request and practically a press release. I'd be surprised if either of them books the flight.

Here's why any of this touches you. Model providers are turning into political entities. Their relationships with governments affect which contracts they win.

Those relationships also affect which countries they operate in, and which jobs they refuse to take. That's a vendor risk, the boring kind, like a supplier with one factory.

So don't build your whole operation on one model provider's API and one prompt style. Not because Anthropic is going anywhere. Because being able to switch is cheap to arrange now and expensive to arrange later.

Why open models now run 56% of Vercel's AI tokens

Vercel, the hosting company a lot of developers use, now runs 56% of its AI tokens on open models. Those are models you download and run yourself instead of calling one provider's API.

That number would have been laughable a year ago. Open weights were the hobbyist option. Now they're the majority at a company with a serious compute bill.

On top of that, a new downloadable model landed this week that you can use commercially. The hosted version costs about 40 cents per million tokens, which undercuts several of the big names. It runs at up to 2,000 tokens per second. That's faster than any human can read.

It also does something the big labs deliberately avoid. Closed providers tend to limit how well their models help you train other models. That's mostly about competition and a little about safety. An open model doesn't care what you use it for.

The reason this shift is happening is simple, and it repeats in every technology. First you pay for quality, because nothing works. Then things work, and you start paying attention to cost. We're in the cost phase now.

Most of what businesses actually automate is not hard. Reading an email and pulling out three fields. Tagging a support ticket. Drafting a reply that a human checks.

That does not need the smartest model on earth. It needs a model that doesn't hallucinate the customer's order number.

So here's the approach I'd take when the bill starts growing. Build it with the best model first, so you know the job is doable at all. Then step down and measure.

• Get the task working with the strongest model you have access to, and save the outputs as your reference.
• Run the same task on the next cheaper model and compare against those outputs, not against your feelings.
• Write down the accuracy drop as a number. One percent, three percent, whatever it is.
• Compare that drop to the cost saving. A one percent drop for a third of the price is often an easy yes.
• Stop stepping down the moment the failures stop being annoying and start being expensive.

The part people skip is deciding the acceptable failure rate before they start. It's different for every process. A mis-tagged newsletter signup costs nothing.

A medical intake form that drops a detail is a different conversation. There the tolerable error rate is zero. A human signs off, full stop.

For most small businesses this optimisation can wait. If you're spending 40 euros a month on tokens, squeezing it to 12 is not the best use of your week. Fix the process first, then the price. When we build AI agents for small teams, the expensive part is almost never the model.

For anyone with a real monthly bill, though, the maths has changed. Cheap open models are no longer a downgrade you apologise for. They're the default for routine work, with an expensive model kept for the hard 10% of cases.

What I'd actually keep an eye on

The two stories are more connected than they look. Politics is making it riskier to depend on a single AI provider. Open models are making it cheaper not to. Those two forces point the same direction.

My guess for next year is that the interesting fights won't be about benchmarks. They'll be about who is allowed to sell what to whom. And about who quietly runs their own model to dodge the question.

Say a government does let an AI approve weapons targets on its own. Put it on record that I wanted independent testing first. A confident answer on a benchmark is no proof that a system can make an irreversible decision.