Open models handle 71.5% of tokens at Vercel

The figure is right

Out of every ten tokens passing through Vercel's AI gateway, seven go to an open model: one whose weights are published and downloadable, like those from DeepSeek, Moonshot (Kimi) or Z.ai (GLM). In September, exactly 71.5%.

The number comes from Vercel itself, in its monthly report published on October 15. Two days later The Register ran it under a headline saying open models are winning hearts, minds and wallets.

We read the whole report, its methodology note, and the three reports that came before it. First finding: the 71.5% is accurate. And Vercel is upfront enough to put the other number in its own headline: those same models account for 23% of spend.

A gateway sees what passes through it

AI Gateway is a Vercel product. An app sends its requests there, and the gateway routes them to hundreds of models at different providers. Think of a switchboard: it knows exactly who is calling whom among its subscribers, and nothing about calls made outside its network.

What escapes the switchboard is considerable. A company that calls OpenAI or Anthropic directly, goes through another gateway, or runs a model on its own servers appears nowhere in these figures. Vercel's customers are developer teams who picked a tool built to swap models easily, and they use it: according to the same report, nearly one team in three changed its main lab in September.

The report doesn't state that limit. Its first sentence says the opposite: this traffic, it claims, gives a view of what AI use in business really looks like today. It's the body of The Register's article that points out the gateway sees neither other gateways, nor direct access, nor on-premises hosting. For once, the write-up is more cautious than its source.

A token is not a dollar

A token is a chunk of text, the unit in which a model reads and writes and in which providers bill. But they don't all charge the same price, far from it.

Counting tokens is like counting cars at a toll booth without checking whether they're compact cars or trucks. On Vercel's gateway, the compact cars are Chinese and cheap.

DeepSeek alone accounts for 58% of all September tokens, for 6.6% of spend. And a single one of its models, DeepSeek-v4.1-Flash, handles nearly half of the month's tokens for 2.6% of the money. Its price per token runs at about a fifth of the open-model average.

At the other end, Moonshot, the lab behind Kimi: 3.6% of tokens, 13.2% of spend. Vercel notes that roughly two thirds of the rise in open-model spend comes from the pricier models in that family, Kimi K3 first among them.

So The Register's headline stretches the point, and its subhead corrects it right away by talking about nearly a quarter of estimated spend. Three quarters of the money, 77% to be exact, still goes to closed models. And that amount is itself an estimate at list prices: the methodology note warns that actual invoices may differ.

One last detail on units: Vercel counts in its tokens those a model rereads from a cache, on top of those it actually reads and writes. It's spelled out plainly, and it's one more reason not to confuse tokens with work done.

The same July, measured twice

The curve starts to move when you go back through the earlier reports.

The report published on August 11 gave the open-model share for July: 36% of tokens. The October report, in its chart, gives 46.5% for the same month of July. For June, the July report said 29%, while October's shows 34.4%. No error and no quiet edit: the rule changed.

Up to the August report, Vercel counted as open only the tokens served by four labs that host their own models (DeepSeek, MiniMax, Moonshot and Z.ai). That was a definition the July report itself called conservative. Since the September report, it's the gateway's model list that decides. The July report already described this other way of counting, by model rather than by provider, for a different indicator. The methodology note flags it in one line: a broader definition than before, and past months that may be revised.

It's a bit like measuring a child every month and switching to a different height chart partway through the year. The growth is real, but the jump in the month of the change partly belongs to the chart.

In its last two reports, Vercel compares its months on a single rule. Not everyone does.

On September 18, The New Stack wrote that the previous report put July at 36%, right after announcing 56% for August. On the 22nd, Techstrong.ai spoke of a big jump from July's 36%, in its text, its summary and its FAQ. On a constant rule, the move from July to August is about nine points (46.5, then 55.8), not twenty.

Both articles are otherwise solid: they explain the gap between tokens and spend, and Techstrong even notes what the gateway can't see. The mistake comes down to one line of a methodology note that nobody picked up.

What still holds

On the current rule, applied to every month, the trend holds. Open models go from 34.4% of tokens in June to 71.5% in September, and from 4.9% of spend to 22.6%. More than quadrupled, as Vercel writes, and this time the calculation uses the same chart throughout.

Another solid finding, and probably the most unexpected in the report: close to 70% of the money spent on Chinese models through the gateway goes to US-based hosts. The Chinese lab supplies the model, an American company runs it and collects the revenue.

What remains is what the figure doesn't measure, and what no write-up can make it say: open models' share of AI in general. A gateway can't give you that number. It only knows what its own customers entrusted to it in September, and it changed the rule for counting it.