Open weight takes 71.5% of tokens and 23% of spend

AI Gateway traffic shows open-weight models now handle roughly 71.5% of tokens while accounting for about 23% of spend, a share that has more than quadrupled since June. Overall gateway spend ballooned - September was ~13× December - while average token price fell ~47%. Token growth is driven by very cheap models (e.g., DeepSeek-v4.1-Flash captured nearly half of September tokens but only a small share of spend), whereas high-priced open-weight models (led by Kimi K3 and Anthropic’s Opus family) are driving most of the dollar growth. Frontier-model spending has surged (roughly tenfold/900%) even as its share of total spend declined, and many teams are diversifying: nearly half switched their most-used model and almost a third switched labs in September. Higher-volume teams use dozens of distinct models (averages: ~15 models for teams with ≥10M tokens, ~26 for ≥100M, ~42 for ≥1B).

American providers are increasingly selling access to Chinese-origin models: in September more than two-thirds of spend on Chinese models flowed through US-based providers, up from under 10% in December. New launches show different success patterns: TypeSafe’s Jev was the fastest-adopted (reaching 35% of paid teams quickly) but represented only a tiny share of spend, while Anthropic’s Opus 5.5 captured ~11% of launch-period dollars and the Opus family took nearly a third of gateway dollars in September. OpenAI’s GPT-6 Luna drew high retention; GPT-6 Astra attracted far more spend per user. The market is rapidly expanding, getting cheaper per token, and becoming more competitive and fluid.