Open-weight models carry 7 in 10 tokens on Vercel

If you still think open models are a side project, Vercel's traffic disagrees.

Vercel's AI Gateway Production Index, published Oct. 15 and covering September, says open-weight models now process more than seven in ten tokens that pass through its gateway. They account for nearly one in four dollars spent there, up from one in twenty in June.

The closed labs still take most of the money. The open ones do most of the work.

DeepSeek moves the tokens, Claude takes the dollars

DeepSeek alone carried 58% of the gateway's September tokens for 6.6% of its spend, Vercel says. Its V4.1 Flash model handled nearly half of all September tokens for 2.6% of spend.

On the money side, Claude Opus 5.5 took 11.03% of spend in its first two weeks, the largest share of any September launch, and GPT-6 Astra took 4.86%. Vercel says Anthropic's overall share of spend fell below half for the first time, to 43.9% from 63.7%.

The fastest-adopted model in the gateway's history is one most people have not heard of. Jev, from TypeSafe AI, launched on the gateway Sept. 17 and reached 35.3% of paid teams in its first two weeks, according to Vercel. Claude Sonnet 5.5 was second on that measure at 18.1%. Vercel calls Jev a "decision model," faster and cheaper than typical requests, and it drew a fraction of 1% of spend.

The market stopped picking one lab

Two numbers show how fast this is moving.

September spend through the gateway was about 13 times December's, while the average price per token fell 47%.

Nearly one in three teams changed their most-used lab in September.

Vercel calls that a record for the index. It is the shape of a market with no loyalty. Teams send bulk, cheap work to open models, keep the hard calls on Claude or GPT, and move the moment a better deal shows up.

A fair caveat: this is Vercel's traffic only, spend is estimated at list prices, and Vercel says actual bills may differ. A gateway full of developers trying models is a slice of the market, not all of it.

Price your own token mix

Pull your last month of usage and split it by task. Classification, extraction and routing calls are where a fast open model earns its place. Reasoning and code you ship are where a frontier model still earns its price.

If you use DeepSeek, Kimi or GLM through an API, China's draft export rules in today's edition leave that kind of access alone. And Reflection's Beam just added a U.S.-built option under Apache 2.0.

Watch the November index

Watch two numbers in Vercel's next report: whether open-weight spend keeps climbing past one in four dollars, and whether Jev keeps its users. Vercel says 55.6% of teams that tried it in week one came back in week two.

Route by task. Then check the bill.