AI models that anyone can download and run are called open-weight models. On one day in mid-September, they handled 78.4% of the traffic on an AI service run by Vercel. About twelve weeks earlier, on June 24, their share had been 28.4%.
Like all AI models, these read and write text in small chunks called tokens. Vercel’s service passes tens of trillions of them a month between companies’ apps and AI labs. Its report on July usage said open-weight models crossed a third of all traffic at about a seventh of the prices the leading labs charge.
Price decides how much money a share of traffic brings in. In July, models from Anthropic, the maker of Claude, handled 30% of the service’s tokens but collected 65% of the spending. Their average price per token was 4.4 times the average for every other lab. Vercel values each request at the lab’s published list price, and the figures cover only its own service. “Cheap open-weight models have been taking volume for months without taking revenue,” the report says.
AT&T shows how the price difference looks to a buyer. It first used closed models, the kind that cannot be downloaded, for customer service, call transcription and coding. As AI costs climbed, its chief data and AI officer, Andy Markus, looked for cheaper options, a New York Times report says. Markus said open-weight models had risen to 40% of AT&T’s AI use. He added that the company was saving up to 80% on AI costs compared with earlier in 2026.
The Times report flags possible trouble for Anthropic and OpenAI, the maker of ChatGPT, as open-weight models grow more popular. Both are heading toward initial public offerings. On June 1, Anthropic said it had confidentially submitted a draft registration statement to the Securities and Exchange Commission. OpenAI followed with a confidential filing, though at the time it had not decided when, or whether, to go ahead with a listing.
Then, on September 22, days after that mid-September reading, both labs released models at lower prices. Anthropic says its Opus 5.5 model needs less computing power than the model before it, so it costs 20% less. The same day, OpenAI released GPT-6 Sol and Luna at half the promotional prices of their GPT-5.6 versions, and credited improvements in how it runs the models.
