The gateway report shows a shift in token consumption and spending within one slice of the AI market.
The Register reports Vercel’s data on its AI Gateway, which routes customers’ requests to many providers. Open-weight models more than quadrupled their share of spending between June and September, when they processed more than seven in 10 tokens (the headline gives 71.5%). In August they processed the majority of gateway tokens for the first time; their share of spend rose from one in twenty gateway dollars in June to nearly one in four in September. Gateway spending in September was about 13 times December 2025’s, while average price per token was about 47% lower, and almost one in three teams changed their most-used provider in September.
By spending, Claude Opus 5.5 took 11.03%, GPT-6 Astra 4.86% and DeepSeek V4.1 Flash 4.19%, ranking third. TypeSafe AI’s Jev was the most popular model, reaching 35% of paid teams in 14 days, but only 1% of tokens and 0.15% of spending. The article notes the gateway does not see usage through other gateways or direct vendor access, so it is one slice of the market.
In plain English
What this means, simply: many teams now run cheap open-weight models for most requests and pay more only for the hardest ones. The concept: routing, sending each request to a model sized for the job. The lesson for you: if you build with AI, design so you can switch models, because price and quality keep moving.
Why it matters
It is a measured shift toward open-weight models in real usage, from one gateway’s data, rather than a launch claim.
Token share, spending share and the number of teams using a model measure different things. A low-cost model can process many requests while receiving a smaller share of total spending. A model tried by many teams may still account for little production traffic. Those differences are visible in the reported gateway figures.
The data therefore supports a narrower conclusion than a declaration that one kind of model has won the whole market. It describes activity passing through one service. Direct purchases, other gateways and workloads outside that service are not represented by the same sample.
For a team buying inference, the useful question is which tasks can be handled acceptably by a less expensive option. A routing experiment can compare simple classification, extraction and summarization separately from difficult reasoning or coding. Define acceptable quality first, then measure cost and failure rates against that standard.
Switching providers also has costs that a token-price chart will not capture. Prompts, tool interfaces and output expectations may need adjustment. Keeping those dependencies visible makes it easier to evaluate a change without assuming models are interchangeable. The report provides evidence of customers moving among options, but it does not establish the best provider for any particular application.