Seven in ten AI tokens now run on open weights and your closed-model reseller is next

Vercel's AI Gateway now sees open-weight models handling 71.5% of the tokens it routes while collecting about 23% of the money. If your inference bill is still dominated by one frontier lab, I think you are paying a premium on work that the rest of the market has already moved somewhere cheaper. If you sell software whose margin is a markup on a closed-model API, I think your customers are about to work that out too.

I expected the open-weight story to stay a volume story: cheap models doing cheap chores while the frontier labs kept the money. That held in June, when open weights took one in twenty gateway dollars. By September they took nearly one in four. Since December, open-weight spend has multiplied more than seventy times over, while frontier spend has grown roughly tenfold.

Look at DeepSeek-v4.1-Flash. It handled close to half the gateway's September tokens and earned 2.6% of spend, charging roughly a fifth of what the average open-weight model costs per token. Vercel doesn't say what those requests were. My guess is classification, extraction, tagging and summarising, the plumbing of every AI product. Somebody is doing that plumbing at Flash prices. If you are doing it on a frontier model, you are paying for capability the task never touches.

The buyers who have noticed are moving fast. In September close to half of teams swapped the model they lean on most, and nearly a third switched their main lab, the highest rate the index has recorded. Teams that pushed at least a billion tokens since December used an average of 42 models. Those are companies with routing in place, sending each job to the cheapest model that can do it. The overpaying buyer is the one with a single API key and a single annual commitment.

The frontier labs have a decent reply. Their revenue is rising. Of the teams that spent more on open weights in September, 68% also spent more on closed models. Among existing frontier customers, the share that raised spending from August passed two in five, its highest point this year. In its first fortnight, Opus 5.5 captured over a tenth of all gateway spend. Buyers clearly still pay up for top-end models.

I accept all of that, and I read it as a lab getting more money from a customer while losing most of the customer's workloads. Gateway spend in September was about 13 times December's, so growing revenue inside that market is easy. Holding share is the hard part. Almost nine in ten Opus 5.5 dollars came from teams already on Opus 5, so the launch mostly recycled existing customers. Across all versions Opus fell from one in two gateway dollars in August to almost a third in September, and Anthropic's share of total spend dropped from 63.7% to 43.9%. In my view frontier money is concentrating on the hard problems, which is where it belongs, and leaving everything else.

The other objection is sharper. Models priced at least four times the open-weight average, led by Kimi K3, drove about two-thirds of the rise in open-weight spend. So the token-to-spend gap partly reflects cheap models doing simple work. I take that as worse news for the closed labs, because open weights are now winning premium jobs too. I would add two caveats Vercel itself flags: spend is estimated at list prices, and the open-weight definition is broader than in earlier reports. I don't think either caveat changes which way the money is moving.

Now the resellers. Hosting open weights is a fine business. US-based providers handled close to 70% of the gateway money spent on Chinese models in September, against under 10% in December. Moonshot took 3.6% of tokens and 13.2% of spend, and American hosts served nine in ten of those Kimi tokens. The vendors I would worry about are the ones whose pricing assumes the customer can't see the cost underneath. The average price per token on the gateway fell 47% while spend rose. A buyer whose own gateway can swap the model under your product will ask why your price didn't move. Popularity won't save you either: Jev reached 35% of paid teams in two weeks and took 0.27% of spend.

So pull last month's inference invoice and tag every request by task. Price everything that isn't long reasoning or hard code generation at an open-weight rate. If most of your frontier spend lands in that pile, put routing in place this quarter and walk into your next renewal with the repriced figure in hand. If you are the vendor on the other side of that renewal, run the same exercise on your customers' usage before they do.

Prompted by Open weight takes 71.5% of tokens and 23% of spend, Vercel.