71.5% of the Tokens, 23% of the Bill: One Gateway's September Ledger

One gateway, one month, one uncomfortable ratio

Vercel published its AI Gateway Production Index on October 15, covering September traffic. The headline number is easy to misread: open-weight models handled 71.5% of the tokens routed through the gateway and took about 23% of the spend.

Read it quickly and it sounds like a victory lap for open models. Read it carefully and it says something more useful. For every dollar spent on a closed model, roughly thirty cents bought eight times as many tokens. The volume and the money have separated, and the separation is the whole story.

The usual caveats come first, because they matter. This is one company's gateway, carrying tens of trillions of tokens between production applications and labs. Enterprise customers of the big labs often call those APIs directly, so a gateway undercounts them. Spend is estimated at list prices, and real bills involve discounts. Vercel also sells the gateway, so it is grading its own business. Treat this as a good sample of what multi-model teams do, not a census of the market.

The cheap model is not a trend, it is one model

The most clarifying line in the report is about DeepSeek-V4.1-Flash: it processed nearly half of all September tokens and accounted for 2.6% of spend. Its price per token was roughly a fifth of the open-weight average.

That single sentence reframes the 71.5%. It is not a broad tide of open models taking over production. A large part of it is one very cheap model doing a very large amount of work. If you are trying to learn a lesson from the index, the lesson is about task routing, not about license ideology.

The money side is equally specific. Open-weight spend share went from about 5% in June to nearly a quarter in September, more than a fourfold climb from a small base. But roughly two-thirds of that increase came from expensive open models, led by Kimi K3, each costing at least four times the open-weight average per token. Cheap models drive volume; expensive models drive spend. That is true inside the open-weight category, not just between open and closed.

Anthropic's share of gateway spend fell from 63.7% to 43.9%, under half for the first time. The tempting reading is that Anthropic lost business. The report says the opposite happened: frontier spend grew tenfold since December. Anthropic is earning more money from a smaller slice of a much larger pie. That is what share loss looks like in a market that is growing 13× while the average price per token falls 47%.

Switching is now a monthly habit

Nearly one in two teams changed their most-used model in September. Almost one in three changed their most-used lab, which the index calls a record. Both numbers have risen for five consecutive months.

That is a strange thing to see in enterprise software. Infrastructure choices usually calcify. Model choices are doing the opposite, and the report offers a mechanism: teams that processed at least a billion tokens since December used an average of 42 different models. Teams between ten and a hundred million tokens averaged 15. Even small teams run a mix.

The reason is boring and correct. Different tasks have wildly different price-quality curves. Classifying a support ticket, extracting fields from an invoice, and writing the customer-facing apology are three different problems that were never going to be solved by the same model at the same price.

The operational implication is less boring. If a model name is hard-coded in thirty places in your application, every price change is a refactor. If it sits behind a routing layer with your own evaluations, it is a config change. Half of your peers are re-deciding their default model every month. The teams that can act on that without a sprint are the ones who capture the savings.

The adoption-versus-revenue split

Two September launches make the point better than any chart.

Jev, a decision model from TypeSafe AI, reached 35.3% of paying teams within two weeks, the fastest adoption in the gateway's history, nearly double the second-place Claude Sonnet 5.5 at 18.1%. Over the same window it took 0.27% of spend. It is a model built for quick decisions: cheap, fast, light on tokens. Plenty of teams tried it. It was never going to be expensive.

Opus 5.5 went the other way: more than one in ten gateway dollars in its first two weeks, the largest spend share of any September launch.

Put together, they describe a two-tier market that has nothing to do with open versus closed. One tier is high-frequency, low-stakes calls where the winner is whoever costs least per correct answer. The other is low-frequency, high-stakes calls where people will pay for the last few points of quality. A vendor can dominate one tier and be invisible in the other, and both can be healthy businesses.

The China detail nobody puts in the headline

Nearly 70% of September's gateway spend on Chinese-developed models went through US-based providers, up from under 10% in December. In rough terms, about $2.50 moved through American providers for every $1 through China-based ones.

This is a routing fact dressed up as a geopolitical one. Enterprises want the price of a Chinese open-weight model and the billing, latency, and data-handling posture of a US vendor. The gateway exists precisely to make that combination possible, and the money followed the combination rather than the flag.

It also means the "open models are winning" narrative gets complicated depending on who you ask. The model weights may be Chinese. The invoice may be American. The team making the decision is probably in neither place.

What to do with this

Three things, and none of them require an opinion about open weights.

Tag every AI call by task. You cannot route what you cannot see. Most teams have a rough sense of which features cost the most and almost no sense of which calls inside those features do.

Price your top three tasks against two models each. One frontier, one cheap. Not on a leaderboard, on twenty real examples with known good answers. Twenty is enough to catch a swap that broke formatting or dropped a field without anyone noticing.

Assume your default model will change within a quarter. Half of the teams in this index already changed theirs last month. Building for that is not over-engineering; it is the current baseline.

The real conclusion is unglamorous. Model choice has become a procurement decision that gets revisited monthly, and the bill follows the hard parts of the work rather than the brand on the API. That is a healthier market than the one we had a year ago, and a much more annoying one to plan for.