Open models pass half of Vercel’s token volume while taking a smaller share of spend

Two measures tell different stories

Open-weight models processed 56% of Vercel AI Gateway tokens in August and represented 14% of estimated spending. The figures come from the September 17 report. They describe traffic through that service, not the entire AI market.

Why this matters for content budgets

An editorial workflow may combine research, drafting, review and images. Evaluating its cost requires defining what each run includes and how many corrections it needs. A low price per token answers only part of that question.

To compare options, we recommend using the same assignment and recording three outcomes:

• Quality: factual errors, unsupported citations and missed constraints.
• Additional work: human revisions, retries and unfinished tasks.
• Total cost: the calls, searches and image generation actually used.

These are criteria for your own trial. Vercel’s index does not measure the cost of producing a HeyMark news article.

Choose for the task

A team can test different models for jobs with different requirements. Keep the same verification standard and measure the complete outcome before changing a stable configuration.

Our article on evaluating AI tools develops that comparison. The useful next question is which model completes your assignment at an acceptable cost after the output has been checked and corrected.