Open-weight models take over half of Vercel's AI Gateway traffic

Open-weight models ran 56% of tokens on Vercel's AI Gateway in August 2026, the first majority — but only about 14% of spend. A separate 78.4% figure is one day, not this index.

In this article

Tokens are not spend
What changed between editions
One figure that is not this index

Open-weight AI models accounted for 56% of the traffic running through Vercel's AI Gateway in August 2026, the first month they took a majority of token volume on the platform, according to Vercel's September 2026 Production Index. That share was below 10% as recently as December 2025, and climbed every month from April through August.

A multi-lane motorway at dusk with long red and white light trails from traffic under a deep blue sky — Photo: Krzysztof Popławski (CC BY 4.0), Wikimedia Commons.

Tokens are not spend

The token-share figure only tells part of the story. Vercel's own data shows open-weight models accounted for about 14% of estimated spend on the gateway over the same period, even while running the majority of tokens — a gap that reflects lower per-token pricing on open-weight models rather than lighter real-world usage. Average price per token on the gateway fell 23.2% in August alone, the third consecutive monthly drop.

What changed between editions

The same report covers a fast launch: OpenAI's GPT-6 Astra, released 3 September 2026, took about 7.7% of all gateway spend in its first twelve days, roughly twice Anthropic's Fable 5.1 at 3.7% over a comparable window, and reached about a third of OpenAI's own spend share within 48 hours. Vercel also noted that Fable 5, Anthropic's flagship model at the time, lost roughly two-thirds of its gateway spend share in the same month.

One figure that is not this index

A separate, frequently cited figure — that open-weight models handled 78.4% of gateway token volume — comes from a single-day snapshot Vercel's CEO shared separately, for 19 September 2026, not from this monthly index. A one-day data point does not establish an ongoing three-quarters market share, and MSPs citing gateway statistics should be clear about which of the two numbers they mean.

All of these figures describe traffic on one production gateway, not the AI market as a whole. They are useful for capacity planning and vendor diversification decisions on that platform, not as a census of who is winning.

Sources: Vercel — AI Gateway Production Index, September 2026, RuntimeWire — on the separate 78.4% daily figure.