Open-weight AI models' token-to-spend split

Gateway data from multiple providers shows open-weight models handling the majority of token volume while generating a small share of revenue. Requesty's data for the week of September 21, 2026 shows open-weight models at 54.6% of tokens, 34.3% of requests, and only 14.9% of spend across the 40 largest models on its platform. Vercel reported 78.4% of its AI Gateway token volume going to open-weight models on September 19, 2026. Earlier Vercel enterprise data showed open-weight models growing from 11% of token traffic in April 2026 to 29% in June 2026, while representing under 4% of AI spending. The token-to-spend gap is partly structural: Requesty notes that cheap models attract the longest, most repetitive agent loops, and open-weight Flash-tier models on its platform run cache hit rates of 81-96%, against a platform-wide rate of 78.3%. Agent workloads reinforce this pattern. A16z data cited by Rohan Paul shows agents burn nearly 5x the tokens humans do, a figure that grew 14x in six months, with child agents in swarms routed to the cheapest model that passes task verification.