Open-weight AI models handled 71.5 per cent of the tokens sent through Vercel's AI Gateway and took nearly a quarter of the estimated spending, according to figures published by the company and reported by Thomas Claburn in The Register.
"In August, open-weight models processed the majority of AI Gateway token volume for the first time," Vercel said. "The money is now following. Open weight's share of spend has more than quadrupled over the last three months, from one in twenty gateway dollars in June to nearly one in four in September."
The gateway is an intermediary, like OpenRouter, that allows customers to purchase tokens from different AI providers through a common interface. It works with popular harnesses and provides access to both proprietary and open-weight models, but it does not see usage through other gateways, or from customers who buy directly from vendors or host models on their own premises. The figures are a slice of the market rather than the market.
Within that slice, the direction of travel is not in doubt. September gateway spending was about 13 times higher than in December 2025, while the average price per token was about 47 per cent lower. Almost one in three teams changed their most-used AI provider in September, which The Register notes suggests that competition may be limiting vendor lock-in.
American models still take the largest share of the money. In September, 11.03 per cent of spending went to Claude Opus 5.5, followed by GPT-6 Astra at 4.86 per cent. DeepSeek V4.1 Flash ranked third, at 4.19 per cent.
The provider breakdown is where the two measures part company. Moonshot AI, maker of Kimi, captured 3.6 per cent of gateway tokens and 13.2 per cent of spend. DeepSeek took 58 per cent of tokens for just 6.6 per cent of spend. Z.ai, maker of GLM-5.3, took 7.2 per cent of tokens and 4.5 per cent of spend.
The month's most popular model was TypeSafe AI's Jev, which reached 35 per cent of paid teams using the gateway in just 14 days. Its popularity did not generate much spending: usage accounted for only 1 per cent of the month's tokens and 0.15 per cent of spending, which the company attributed to the model's architecture.
More than two-thirds of September's spending on Chinese models went to US providers serving those models, up from less than a tenth in December 2025, Vercel said.
A market that reports the same activity in two currencies will sooner or later have to say which one it means. Measured in tokens, DeepSeek is most of the way to owning the gateway. Measured in money, it is a rounding error with a following.
The two readings can be reconciled, but the exchange rate is brutal and it runs in both directions. Divide spend by tokens and DeepSeek's position shrinks by a factor of nearly nine. Do the same for Moonshot and its position grows by a factor of nearly four. Market leadership is accordingly a question of which column the leader is reading from.
September's most popular model demonstrates the principle at the limit. Jev reached 35 per cent of paying teams in a fortnight, which by adoption makes it the product of the month, and collected 0.15 per cent of the spending, which by revenue makes it almost nothing at all. Both figures describe the same fortnight.
Our analysts have extended the two trend lines that Vercel supplies. Volume up thirteenfold, average price per token down 47 per cent: continue both and the gateway converges on a market that processes everything and is paid for none of it. At that point share of tokens and share of spend finally agree, at 100 per cent and nothing respectively.
The Department of Unnecessary Research observes that this is already the recorded position of a model that a third of paying teams adopted in a fortnight, and that it is indistinguishable, in the spending column, from the position of a model nobody has adopted at all. The distinction is maintained entirely by the billing department.
On these figures the surest route to market leadership is to be used by everybody and invoiced to nobody. Open weights are 71.5 per cent of the way there.