News

Open-weight models pass half of tokens in Vercel's AI Gateway

Vercel says open-weight models processed 56% of tokens in August but represented 14% of spending. What the figures show, and what cannot be inferred from them.

Open-weight models processed 56% of tokens passing through Vercel's AI Gateway in August, compared with 7% in December 2025. Despite their majority in volume, their share of spending was just 14%. The figures appear in the company's September report, whose page displays a September 17 date and covers traffic through August.

This is a figure about Vercel's service gateway, not the entire artificial intelligence market. The distinction narrows the conclusion's scope but leaves an interesting business question: how should tasks be divided between models when usage volume and spending tell different stories?

A majority in volume is not a majority in business value

According to Vercel, this is the first time open-weight models have passed half of token volume in the gateway. Tokens are text-processing units; counting them is not equivalent to counting users or successfully completed tasks.

The gap between 56% of volume and 14% of spending also does not prove that every switch to an open model will make a particular task cheaper. Price, answer length, number of attempts and type of work affect the bill. It is inaccurate to infer a guaranteed saving rate for a new customer from these two percentages.

The useful decision is task-specific

When assessing business suitability, start with a defined set of tasks, such as classifying inquiries or extracting fields from documents, and compare outputs against an agreed quality measure. Only then calculate the cost of an accepted result, including failed attempts and human correction time.

Routing between models can be a way to make this division: one task goes to a particular model, another to a model with different capabilities or pricing. Vercel's figures do not prove this is the right solution for every company, but they give a reason to test the assumption that all activity should pass through the same model.

Opening the weights does not necessarily open everything

The term open weights refers to the availability of model parameters. It alone does not mean training data is public, the license is unrestricted or commercial use is permitted under all conditions. Read the particular model's license before choosing it.

There is also a difference between consuming a model through an API and self-hosting it. Self-hosting brings computing resources, maintenance and operations into the calculation. For an Israeli business, a good test will also include Hebrew examples from real work. Price per token is a starting point for comparison; the cost of a useful result is the measure to finish with.