Open-weight models: 71.5% of tokens and 23% of spend on the Vercel AI Gateway

On 15 October 2026 Vercel published the AI Gateway Production Index for September, and on 19 October heise picked it up. The headline: open-weight models processed 71.5 percent of all gateway tokens in September, more than seven in ten. Their share of spend is 23 percent.

Source: Vercel AI Gateway Production Index, 15 October 2026, traffic through September 2026. This is Vercel's own gateway data, not the whole market. (Quelle: Simple Icons, file vercel.svg (CC0))

The numbers

Three months earlier the open-weight share of spend was about one dollar in twenty, and Vercel calls the rise more than a quadrupling. Since December 2025 spend on open models has grown more than seventyfold, Vercel says, and spend on proprietary frontier models tenfold. Total gateway spend in September was about 13 times December's, and the average price per token fell 47 percent.

The second finding concerns Anthropic. Its share of spend dropped from 63.7 percent in August to 43.9 percent in September. All Opus versions together account for just under a third of September dollars, down from about half in August.

The June value is Vercel's wording of about one dollar in twenty. Spend is estimated from published list prices. (Quelle: Vercel, AI Gateway Production Index, 15 Oct 2026)

Why token share and spend share sit so far apart

The gap between 71.5 and 23 percent comes from prices. DeepSeek alone is 58 percent of tokens and 6.6 percent of spend. Moonshot AI with Kimi is the mirror image: 3.6 percent of tokens but 13.2 percent of spend. Z.ai (GLM) has 7.2 percent of tokens and 4.5 percent of spend. Reading "open" as a synonym for "cheap" is not always right.

Where open Chinese models run is worth a look. Nearly 70 percent of spend on Chinese-origin models went through US-based providers, against under 10 percent in December. For Moonshot, US providers served nine in ten tokens, for Z.ai seven in ten, while for DeepSeek the vendor itself served about four in five.

What the numbers do not say

Vercel names the limits itself. The data is anonymised traffic through its own gateway. Spend is estimated from list prices, and real bills can differ. The open-weight classification follows the current model list, which is broader than in earlier reports, and earlier months can be revised. heise adds that token share and spend share are not interchangeable, since token prices vary widely.

There is also a bias neither text quantifies: teams that use a gateway like this want to switch between models anyway. Nearly one team in two (48 percent per heise) changed its most-used model in September. A team that runs a model locally does not appear in this data at all.

What it means for local models

In one respect the numbers support the case for open weights: open models have arrived in everyday developer work and carry most of the tokens processed. For local operation they say little, because they measure hosted use, a large part of it through US providers. If you want to keep data in-house, the statistic adds nothing to what privacy and cost already tell you.

Our own lanes are unaffected. We run open weights on our own cards behind a LiteLLM gateway, and we have not measured how our token mix compares with Vercel's. If you want to hold your own choices against these numbers, pull the same figure from your own gateway: tokens and estimated spend per model, split into open and closed weights.

heise names one more risk. On 16 October China put forward a draft of export controls covering 23 AI technologies, among them models of two trillion parameters or more. The consultation runs until 15 November 2026. API access and already published open technology are exempt in principle, but experts warn that authorities could intervene in licence terms. Whether the draft stays in this form will not be clear before the consultation closes on 15 November 2026.