Open weight takes 71.5% of tokens and 23% of spend

Vercel's monthly AI Gateway report, covering traffic through September, shows open-weight models handling more than seven in ten tokens but taking only about one in four dollars (Vercel). The gap exists because different models produce the volume and the money. DeepSeek's V4.1-Flash alone processed nearly half of all September tokens but accounted for just 2.6% of spend, since its price per token was about a fifth of the open-weight average. Meanwhile, models priced at least four times that average, led by Moonshot AI's Kimi K3, produced about two-thirds of the growth in open-weight spending. Cheap models add tokens; expensive ones, open or closed, add dollars.

The shift in spending does not mean customers are cutting back on frontier labs. Open-weight models went from about one in twenty gateway dollars in June to nearly one in four in September. Frontier spend has still grown roughly tenfold since December, while open-weight spend grew more than seventyfold, and 68% of teams that spent more on open-weight models in September also spent more on closed ones. Total gateway spend in September was about 13 times December's even as the average price per token fell 47%, and almost one in three teams changed the lab they used most. Over the same stretch, Anthropic’s share of gateway spend fell from 63.7% in August to 43.9% in September, below half for the first time.

Two measurement notes matter when comparing months. Spending is estimated from published list prices, so actual bills may differ, and this report counts models as open-weight using the gateway's current model list, which Vercel says is broader than the definition in earlier reports. Vercel also notes that prior months may be revised as its methodology changes, so figures from earlier reports, such as last month's 56% of August tokens and 14% of spend for open-weight models, may not line up exactly with this one (Vercel, About this report).