Vercel AI spending soars as customers switch models

Spending on Vercel's AI Gateway has risen 13-fold since December, while the average cost per token has fallen 47% over the same period.

The figures come from Vercel's latest AI Gateway Production Index, which tracks anonymised aggregate traffic across its AI model routing service. The data points to a market in which businesses are using more AI while switching models and providers more often.

Open-weight models accounted for 71.5% of all tokens processed on the gateway in September, up from 56% in August. They made up 23% of spending, more than four times their share in June.

The shift suggests lower-cost, openly available models are taking a larger role in production workloads, even as spending on closed or frontier systems continues to rise. Since December, spending on frontier models has increased tenfold, while spending on open-weight models has grown more than seventyfold.

The pattern also reflects a split within the open-weight market itself. DeepSeek-v4.1-Flash processed nearly half of all September tokens on the gateway but accounted for just 2.6% of spend, while higher-priced models such as Kimi K3 contributed a larger share of open-weight revenue.

Switching rises

Vercel's data also showed high churn in model selection. Nearly one in three teams changed their most-used AI lab in September, while almost one in two switched their most-used model.

Both measures rose for a fifth consecutive month. Among paid teams that processed at least 10 million tokens between December and September, customers used an average of 15 different models. That rose to 26 for teams that processed at least 100 million tokens and 42 for those that processed at least 1 billion.

Those figures indicate that many businesses are spreading workloads across a wide range of AI systems rather than relying on a single supplier. For providers, that means revenue can grow even as customer dependence weakens and market share shifts more quickly.

Chinese models

Another notable change was in how customers accessed Chinese-developed models. Nearly 70% of September spending on Chinese-origin models went through US-based providers, up from less than 10% in December.

In effect, customers are increasingly buying access to Chinese models from US intermediaries rather than directly from the original labs. The report said more than USD $2.50 was spent on Chinese models through US-based providers for every USD $1 spent through China-based providers.

DeepSeek's models processed 58% of all gateway tokens but represented 6.6% of spending. 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, the Beijing-based developer behind Kimi, accounted for 3.6% of gateway tokens and 13.2% of spend in September, with US-based providers serving nine in 10 of those tokens. Z.ai, creator of GLM, represented 7.2% of all tokens and 4.5% of spend, with US providers serving seven in 10.

Launch patterns

The index also tracked new model launches. Jev was the fastest-adopted model in AI Gateway history, reaching 35% of paid teams in its first 14 days.

Despite that uptake, Jev accounted for only 0.27% of gateway spending over the same period in one summary measure, and Vercel's broader September analysis said it represented a very small share of total monthly spend. It described Jev as a decision model that uses fewer tokens and costs less per request than a typical language model.

By contrast, Opus 5.5 drew the largest share of spending among September launches. In its first two weeks, it accounted for more than one in 10 gateway dollars, ahead of GPT-6 Astra and DeepSeek-v4.1-Flash.

The spending data also highlighted a shift in Anthropic's position on the gateway. Across all versions, Opus represented almost a third of gateway dollars in September, down from one in two in August, while Anthropic's overall spend share fell from 63.7% to 43.9%.

OpenAI's launches showed a split between reach and revenue. GPT-6 Luna reached a broader set of teams and posted the highest week-two return rate of the month, while GPT-6 Astra drew a smaller audience but captured more budget.

Vercel said AI Gateway routes tens of trillions of tokens each month between production applications and AI labs. The latest data suggests that while AI spending is climbing quickly, the market is becoming cheaper, more fragmented and more fluid, with customers moving workloads between labs, models and intermediaries at a record pace.