Open-weight models handle 71.5% of Vercel AI Gateway tokens and 23% of spend

What Vercel published

Vercel has published the October edition of its AI Gateway Production Index, a monthly report built from anonymised, aggregate traffic that runs through its AI Gateway, the service that lets developers call models from many labs through one API. The report, dated 15 October 2026 and written by Amelia Charles and Harpreet Arora, covers usage through the end of September.

Its headline finding is that open-weight models, the ones whose weights anyone can download and host, now process more than seven in ten gateway tokens, 71.5% by the report's title, and account for about 23% of estimated spending. In June that spending share was about one dollar in twenty. Vercel says open weight's share of spend has more than quadrupled in three months, and that September brought the largest monthly gains in both token share and spend share in the index's history.

The figures drew wider attention after The Register wrote them up on 17 October, a day after China proposed export controls that could reach some of the same models. Vercel has used this gateway for other open-model pushes too, such as free access to NVIDIA's Nemotron 3.5 Lightning earlier this week.

Open-weight models now carry most of the traffic

Vercel dates the turn to August, when open-weight models processed the majority of gateway token volume for the first time. September extended that lead, and this time money followed. "The money is now following," the report says.

The growth in spending is happening on both sides of the market rather than as a straight swap. Since December, Vercel says, spending on frontier models through the gateway has grown about tenfold, while spending on open-weight models has grown more than seventyfold. In September, 68% of teams that put more money into open-weight models also increased their spending on closed models, and more than two in five teams already paying for frontier models spent more on them than in August, the highest share this year.

The report's reading is that a lab can earn more from a customer while handling a smaller part of that customer's work. Frontier spending is still rising; spending on alternatives is rising faster.

Cheap tokens, expensive tokens

The gap between token share and spend share comes from price. DeepSeek-V4.1-Flash alone processed nearly half of all September tokens but took just 2.6% of spend, because its price per token was about a fifth of the open-weight average. At the other end, models priced at least four times the open-weight average, led by Moonshot AI's Kimi K3, accounted for about two-thirds of the increase in open-weight spending. A chart in the report says Kimi K3 alone accounted for nearly half of that increase while processing far fewer tokens than DeepSeek's model.

That pattern mirrors the whole gateway. The cheapest models drive token growth and the expensive ones drive spend. It also explains why DeepSeek's latest checkpoint matters to developers watching cost per task; we covered the newest version in our report on DeepSeek-V4.1-Flash-1017.

Overall, September gateway spending was about 13 times December's, an increase of 1,180%, while the average price per token fell 47%. The median paid customer active in both months more than doubled its spend.

Chinese models, American hosts

One of the more striking sections concerns where Chinese models are actually served. More than two-thirds of September's gateway spend on Chinese-origin models, 69.5% by the report's chart, went to US-based providers that host those models, up from less than a tenth in December. Vercel counts a provider's location by its company headquarters. By its arithmetic, about $2.50 went to Chinese models through US-based providers for every $1 spent through China-based ones.

The split varies by lab. DeepSeek's models processed 58% of all gateway tokens for 6.6% of spend; more than four in five of those tokens were served by DeepSeek itself, yet two in five dollars spent on its models went through US-based providers. Moonshot AI accounted for 3.6% of tokens and 13.2% of spend, with US-based providers serving nine in ten of those tokens. Z.ai, the maker of GLM, had 7.2% of tokens and 4.5% of spend, with US providers serving seven in ten.

That matters for policy. The Register noted that China's proposed controls, which we covered in our report on the 23 AI technologies, could in principle reach models such as Kimi K3. It quoted Chris McGuire of the Council on Foreign Relations saying the move "introduces significant uncertainty for US businesses that host or rely on Chinese models." The index shows how much US hosting revenue now depends on those models, at least inside this one gateway.

Teams keep switching models

The report also tracks how often customers change their main model. Among paid teams with at least 10 million paid tokens in both August and September, 48.0% changed their most-used model in September and 28.6% changed their most-used lab, which Vercel calls an index record. Both rates rose for a fifth consecutive month.

Teams also use a wide spread of models. Since December, teams that processed at least 10 million tokens used an average of 15 different models; the figure rose to 26 for teams above 100 million tokens and 42 for those above a billion. Vercel's conclusion is that the model is becoming a replaceable part of a larger investment, so companies can spend more on AI without becoming more dependent on one lab.

September's launch scorecard

Vercel ranked September's launches by reach, spend and return rate over each model's first 14 days. TypeSafe AI's Jev, a decision model rather than a chat model, reached 35.3% of paid teams in its first two weeks, the fastest adoption in the gateway's history and nearly double Claude Sonnet 5.5 in second place. Because a Jev request is short and cheap, it accounted for only about 1% of September tokens and 0.15% of spend. We covered the company behind it in our story on TypeSafe's funding.

Claude Opus 5.5 took the most money: 11.03% of gateway spend in its first two weeks, more than twice any other September launch, followed by GPT-6 Astra at 4.86% and DeepSeek-V4.1-Flash at 4.19%. Almost nine in ten Opus 5.5 dollars came from teams already using Opus 5. Even so, Anthropic's overall share of gateway spend fell below half for the first time, from 63.7% to 43.9%, with the Opus family dropping from about one in two gateway dollars in August to almost a third.

OpenAI's GPT-6 Luna had the best retention of the month: 64.8% of teams that tried it in week one came back in week two. Astra reached fewer teams but captured more than five times Luna's share of spend.

Why this matters for builders and policy

For developers, the practical message is that routing by task is now normal. Cheap open-weight models are handling bulk traffic, while expensive frontier and high-end open models take the hard requests. A gateway or router that makes switching easy is part of why nearly half of teams changed their main model in a month.

For policymakers, the data complicate the idea that Chinese open models are a separate market. In this sample, US companies are earning most of the hosting revenue from them, so restrictions in either direction would hit American providers and their customers too.

What the index cannot tell you

The index has clear limits, and Vercel spells several out. Spending is estimated at published list prices, so actual bills may differ. Token volume includes input, output, reasoning and cached tokens, which favours long agent runs. The open-weight category now follows the gateway's current model list, a broader definition than in earlier reports, and prior months may be revised.

Most important, this is one gateway. It does not see traffic that goes straight to OpenAI, Anthropic or Google, through other routers such as OpenRouter, or to models companies host themselves. Vercel's customers lean towards web developers and start-ups. The numbers are a useful view of production usage, but not a measure of the whole AI market.