Analysis
Open-weight AI models processed more than 7 in 10 tokens -- 71.5% -- on Vercel's AI Gateway in September 2026, up from a minority share in June, according to The Register. Vercel's data shows open-weight models crossed 50% of token volume for the first time in August and kept climbing from there.
The spending shift is even sharper than the usage shift: open-weight models captured roughly 5% of gateway dollars in June and nearly 24% by September, even as the average price per token fell about 47% from December 2025 levels. Gateway spending overall was roughly 13 times higher in September than in December 2025. DeepSeek alone accounted for 58% of tokens and 6.6% of spending, while Moonshot's Kimi took 3.6% of tokens but 13.2% of spending, and Z.ai's GLM-5.3 held 7.2% of tokens and 4.5% of spending -- DeepSeek shipped an MIT-licensed update, V4.1-Flash-1017, the same week.
“Gateway spending overall was roughly 13 times higher in September than in December 2025.”
Nearly a third of teams changed their most-used AI provider in September alone, a churn rate that undercuts the idea that foundation-model relationships carry heavy lock-in. Anthropic's Claude Opus 5.5 still commanded the largest single spending share at 11.03%, with OpenAI's GPT-6 Astra at 4.86%, but closed models' combined share is shrinking against a widening field of cheaper open-weight alternatives. More than two-thirds of September's spending on Chinese models flowed through US-based hosts, up from under a tenth in December 2025, a detail that matters given China's own draft export curbs on some of those same models.