Open-weight models, the kind whose trained files anyone can download and run, handled 71.5% of the tokens (the small chunks of text AI models read and write) that passed through Vercel’s AI Gateway in September, but only 23% of the dollars. Vercel, the web-hosting company, reported the figures in its October 2026 AI Gateway Production Index, a monthly look at traffic on its own product. Two caveats come before the numbers. This is one platform’s traffic, a gateway where developers, many of them building web applications on Vercel, route requests to AI labs through a single connection. It is not a census of the AI market. And Vercel sells the gateway, so it is grading its own business.
The arithmetic is the useful part. If open-weight models take 71.5% of tokens and 23% of spend, closed models take 28.5% of tokens and 77% of spend. Divide one by the other and a closed-model token costs roughly eight times an open-weight one on average. That is close to an order of magnitude, though not quite one, and it is a blended figure across dozens of models. Vercel estimates spend at published list prices, so actual bills, with discounts and negotiated rates, will differ.
The report also says where the volume sits. One low-priced open-weight model, DeepSeek-V4.1-Flash, accounted for close to half the tokens in September yet only 2.6% of spend. Per token, it cost roughly 20% of what open-weight models charge on average. So the 71.5% is not a broad open-weight tide. One cheap model supplies a large part of it. Vercel does not say what jobs those tokens do, so the idea that open models are handling the low-value work is an inference from price, not something the index measures.
The money story is more mixed than the headline split suggests. Since June, open-weight models have gone from about 5% of gateway dollars to nearly 25%, more than a fourfold climb. The starting base was small. And roughly two-thirds of the growth in open-weight spend came from expensive open models, with Kimi K3 in front, each costing per token a minimum of four times what the typical open-weight model does. Vercel also notes that its definition of open-weight now follows the current gateway model list and is broader than in earlier reports, so comparisons across months deserve some caution.
The cleanest illustration of the gap between attention and revenue is Jev, a model from TypeSafe AI. In its first fortnight, 35.3% of paying teams used it, nearly double second-place Claude Sonnet 5.5 at 18.1%, which Vercel calls the fastest adoption in the gateway’s history. Over the same window it took 0.27% of spend. The report offers a plain reason: Jev is a model built for quick decisions. A typical language-model call takes longer and burns more tokens, while a Jev request is cheaper, quicker and light on tokens. Plenty of teams tried it, and it was never going to be expensive. A launch that wins on adoption can still be a rounding error on the bill. (The report’s own text gives 0.15% for Jev’s share across the whole month, a different window from the two-week figure.)
The switching numbers matter most to anyone choosing a supplier. Nearly half of teams swapped the single model they lean on most in September, and almost one in three went further and moved to a different lab as their top choice. The index had not recorded a higher lab-switching rate before. “Most-used” here is ranked by paid tokens, and only teams that bought 10 million or more of them in each of the two months being compared are counted. That is a narrower group than every customer. Within it, a team’s top spot can flip because one high-volume workflow moved, not because the whole company abandoned a vendor. Still, the pattern fits other figures in the report: teams processing at least a billion tokens averaged 42 distinct models in use between December and September.
The token totals add together cache-creation, cached-input, reasoning, output and input tokens, and the report does not break out batch traffic. Counting every kind of token in one total means the volume share and the dollar share are not measuring the same thing, and the report does not say how much of the volume is cached.
For a team deciding where to spend, the gateway data points to a split that is easy to test: send the highest-volume, most repetitive work to a cheap open-weight model, keep a closed model on the tasks you would pay for, and make sure your code can swap either one. Vercel’s numbers suggest plenty of its customers already do.
Based on Vercel’s AI Gateway Production Index, October 2026, published on Vercel’s blog.
