Open-Weight Model Share of AI Gateway Requests

## What this measures

The percentage of all requests routed through Vercel's AI Gateway on a given day that went to open-weight models: models whose weights anyone can download and run, such as DeepSeek, Qwen, GLM, Kimi and Llama. The rest goes to closed models served only through their maker's API. Vercel classifies each model.

## Why it matters

Open-weight models set the price floor for AI. When developers move work to them, closed-model makers lose pricing power on the volume that moves. It is the most direct evidence we have of AI margin compression reaching real workloads.

## How to read it

Higher means more of the gateway runs on open weights. On tokens it rose from about 2.5% on 1 October 2025 to about 78% in late September 2026. Compare tokens with spend: open weights take far more of the volume than of the money, because they are cheap.

## What it does not say

It is the traffic of Vercel's AI Gateway customers, not the whole AI market. Vercel publishes shares, never volumes, so a rising share can mean one lab grew or the others shrank. A single large customer can move a day's share by several points; read the trend over weeks, not one day. Direct API traffic to the labs, and other gateways, are not included. "Open" describes the weights, not who served the model.

## Source

Vercel AI Gateway Leaderboard (https://vercel.com/ai-gateway/leaderboards), daily, from 1 October 2025. © 2026 Vercel. "AI Gateway Leaderboard Data" is licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Palanor stores each day's shares and derives the series shown here; the current, incomplete UTC day is never published.