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

Open models have taken hold in production by dollars as well as by volume. According to Vercel's October Production Index, open-weight models process 71.5% of tokens in AI Gateway and account for about 23% of spend. In June they made up one dollar in twenty; in September it was almost one in four. In August they overtook closed models by token count for the first time, and now the money is following.

Vercel sees tens of trillions of tokens a month flowing between production applications and labs. That is a sample of one platform's customers, not the whole market. But the data comes from real payment flows, not surveys.

Cheap models drive tokens, expensive ones drive revenue

Open-weight growth is coming from both ends of the price scale. Cheap models supply the volume; expensive ones supply the money. DeepSeek-V4.1-Flash alone processed almost half of all September tokens but earned only 2.6% of spend. Its token price is roughly five times below the open-weight average.

The growth in spend looks different. Models whose tokens cost at least four times the open-weight average accounted for about two-thirds of that growth. Kimi K3 leads here: almost half of the increase comes from it, even though it processed noticeably fewer tokens than DeepSeek-V4.1-Flash.

Key figures from the September index:

• open-weight: 71.5% of tokens, about 23% of spend; the share of spend has more than quadrupled since June;
• total spend in September was 13 times the December level (+1180%), and the average token price is 47% lower;
• spend on frontier models has grown 10x since December, and on open-weight models more than 70x;
• 68% of teams that increased their open-weight budget also increased spend on closed models;
• more than 40% of teams paying for frontier models spent more on them than in August. That is the highest figure this year.

Frontier labs are not losing in dollar terms. They are losing share of the customer's budget. A company can pay a lab more and still hand it a smaller share of its workloads. Opus 5.5 took more than a tenth of gateway dollars in its first two weeks, the best result among September launches. The Jev model became the fastest-adopted in the gateway's history: 35% of paying teams within two weeks. But it accounted for only 0.27% of spend. Many tried it; few switched to it.

Lab loyalty is falling

Nearly half of teams changed their most-used model in September. Almost a third changed their most-used lab, a record for the index. Switching has risen for the fifth month in a row.

The model zoo grows with volume. Teams that have processed at least 10M tokens since December use an average of 15 models. At 100M tokens the average is 26, and at a billion it is already 42. The model is becoming a swappable part inside a larger investment: prompts, evals, scaffolding and data stay, while the engine underneath changes.

Chinese models, American bills

A separate thread in the index concerns geography. Almost 70% of September spend on models of Chinese origin went through US-based providers. In December it was under 10%. A Chinese lab builds the model, and an American provider runs it and sells access.

DeepSeek: 58% of all gateway tokens, 6.6% of spend. DeepSeek itself served more than four in five tokens, but two of every five dollars spent on its models went to American providers. Moonshot AI, the maker of Kimi, shows a different picture: 3.6% of tokens and 13.2% of spend, with nine in ten tokens served by US providers.

What to change in your architecture

If your product relies on one model from one vendor, that is already a risk, not a simplification. A practical setup: a cheap open-weight model handles high-volume tasks (classification, extraction, drafts), while an expensive closed model or a large open-weight model handles the hard steps. Routing like this requires your own eval set that you can run on every new model in hours, not weeks.

Second: budget for a falling token price and rising volume at the same time. Vercel's data shows that spend is growing faster than unit cost is falling. Third: for models from Chinese labs, weigh not only quality but also the provider's jurisdiction. The same weights can come from different places with different data terms and SLAs.