Vercel index: open-weight models handle 71.5% of AI tokens but 23% of spend

Open weight owns the tokens, not the dollars

Open-weight models now process 71.5% of the tokens routed through Vercel's AI Gateway while taking just 23% of customer spend, according to Vercel's AI Gateway Production Index for October 2026. The monthly report draws on anonymized, aggregate traffic through September 2026; Vercel says the gateway routes tens of trillions of tokens each month between production applications and AI labs.

Open-weight models crossed the majority of gateway token volume for the first time in August, Vercel reports, and the money is now following: their share of spend has more than quadrupled in three months, from roughly one in twenty gateway dollars in June to nearly one in four in September.

The growth is coming from both ends of the price range. Models priced at least four times the open-weight average, led by Kimi K3, accounted for about two-thirds of the increase in open-weight spend. At the opposite end, DeepSeek-v4.1-Flash alone processed nearly half of all September tokens but took just 2.6% of spend, at a price per token about a fifth of the open-weight average.

The shift is not zero-sum. Since December, spend on frontier models has grown tenfold while spend on open-weight models has grown more than seventyfold. In September, 68% of teams that increased open-weight spending also increased their spending on closed-weight models. Vercel's conclusion: a company can spend more with a frontier lab while relying on it for fewer workloads, so the lab earns more money but a smaller share of the customer's business.

A market 13 times bigger at half the price

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

Loyalty is thin. Nearly one in two teams changed their most-used model in September, and almost one in three changed their most-used lab, an index record. Larger users hedge widely: teams that processed at least 10 million tokens since December used an average of 15 different models, rising to 26 for teams past 100 million tokens and 42 for teams past a billion.

Chinese models are big business for American providers

Nearly 70% of September's gateway spend on Chinese-origin models went through US-based providers, up from less than 10% in December. Vercel summarizes the arrangement bluntly: the Chinese lab builds the model; the American provider runs it and sells access.

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, the Beijing lab behind Kimi, accounted for 3.6% of tokens and 13.2% of spend, with US-based providers serving nine in ten of those tokens. Z.ai, creator of GLM, took 7.2% of tokens and 4.5% of spend, seven in ten of them via US providers.

September launches split adoption from revenue

The report's launch scorecard shows reach and revenue diverging. Jev, described by Vercel as a decision model whose requests are faster, cheaper and use a fraction of the tokens of a typical language-model request, reached 35% of paid teams in its first two weeks — the fastest adoption in gateway history — and received the most requests of any September model despite launching mid-month. It still accounted for only 0.27% of gateway spend over those two weeks, and 55.6% of week-one teams returned in week two, the fifth-highest retention rate.

Opus 5.5 did the reverse, taking more than one in ten gateway dollars in its first two weeks — the largest spend share of any September launch — with almost nine in ten of those dollars coming from teams already using Opus 5. Even so, the Opus family's overall share fell to almost a third of gateway dollars, down from one in two in August, and Anthropic's share of total gateway spend dropped below half for the first time, from 63.7% to 43.9%.

OpenAI's two GPT-6 launches split the difference. GPT-6 Luna reached nearly three times as many teams as GPT-6 Astra and posted the month's best return rate, with almost two-thirds of week-one teams coming back, but it captured less than one in a hundred gateway dollars while Astra took about one in twenty.

Why it matters

The Production Index is one of the few public windows into real production AI traffic rather than benchmarks or vendor announcements. Its headline split — 71.5% of tokens against 23% of spend — quantifies how far inference prices have fallen and how quickly teams treat models as interchangeable parts, increasing total AI spending while reducing dependence on any single lab. It also shows that revenue increasingly accrues to whoever operates inference rather than whoever trains the model, which reshapes who profits from open-weight growth.