DeepSeek Reaches 22.6% of AI Gateway Tokens, Open-Weight Models Hit 29% — Vercel Report

In July 2026, Vercel published its AI Gateway Production Index report for June 2026: open-weight models grew to 29% of all tokens on less than 4% of spend, DeepSeek moved into third place with 22.6% of volume, and Anthropic took 61% of all platform spend.

Why open-weight models surged

Open-weight models processed 29% of all AI Gateway tokens in June 2026, up from 11% in April. That is nearly a threefold increase in two months. They accounted for less than 4% of spend: almost a third of traffic for one twenty-fifth of the budget. According to Vercel, the main contributor was DeepSeek, now the platform's third-largest source of tokens with a 22.6% share, trailing Google by less than two points.

• The report was published in July 2026; the data was collected in June 2026
• Token volume grew 29% month over month, spend grew 27%
• Open-weight: 29% of tokens on <4% of spend (up from 11% in April)
• DeepSeek: 22.6% of volume, third place; Google's share fell to 24%
• Roughly one in eight enterprise customers runs an open-weight model in production

A newcomer is accelerating the trend. Z.ai released GLM 5.2, an open-weight model under the MIT license built for long agentic tasks, priced at roughly one-fifth the cost of Opus 4.8. From the API launch on June 16, 2026 to the end of the month, its daily token volume grew roughly 50-fold, and the model climbed to 11th place on the platform by tokens in the final week.

Who dominates spending

Anthropic took 61% of all AI Gateway spend on 32% of tokens, and kept 72% or more of spend in every high-stakes use case. This illustrates the routing discipline the report documents: bulk work goes to cheap models, while critical work stays on frontier models.

"High-volume work goes to low-cost models, while high-stakes work stays on the frontier," the

Vercel AI Gateway report says.

The split also shows up across modalities. OpenAI's GPT Image generated 53% of images, ahead of Google's Nano Banana (39%). In video, Chinese labs took two-thirds of spend, and xAI's Grok Imagine produced 42% of all clips, more than any other model.

What about price per token

The average price per token stayed flat in June 2026, after rising almost 20% in May. Volume and spend grew at nearly the same rate (29% and 27%), so the average cost per token did not move.

Two opposing forces were at work. Cheap open-weight traffic pulled the average price down, while frontier models became about 12% more expensive per token. The two effects offset each other. One more detail from the month: Anthropic's Claude Fable 5 launched on June 9, 2026 and reached 22% of Opus 4.8's request volume in four days, but a US export control directive suspended access to it for the rest of the month.

What this means

Investment in AI keeps growing fast, but companies are spending more wisely: cheap open-weight models take on the volume, while expensive frontier models handle the high-stakes tasks. An open-weight lab may soon become the second-largest source of tokens on AI Gateway.

Frequently asked questions

What are open-weight models?

These are models with publicly available weights that a company can run on its own infrastructure, such as DeepSeek or Z.ai's GLM 5.2. According to Vercel, in June 2026 they generated 29% of the platform's tokens on just 4% of spend.

Which model leads by token volume?

By volume, Anthropic and Google hold the first and second places, and DeepSeek is third with a 22.6% share, less than two points behind Google, whose share fell to 24%.

What is GLM 5.2?

GLM 5.2 is an open-weight model from Z.ai under the MIT license, built for long agentic tasks and roughly five times cheaper than Opus 4.8. After the API launched on June 16, 2026, its daily token volume grew roughly 50-fold in two weeks.