Eight of the 10 largest models by token volume on OpenRouter are now open-weight, reflecting a shift in AI deployment toward more accessible architectures. A Mozilla report finds that these models trail closed frontier systems by about 4 months, with Kimi K3 nearly matching Sol on Terminal-Bench and GLM-5.2 approaching Claude Opus performance at less than 20% of the cost. Closed models maintain a lead on the most difficult tasks, though the performance gap has diminished.

Token volume for open models hit a record 78.4% on the Vercel AI Gateway on Sept. 20, while closed models accounted for 21.6%. Some developers argue that these percentages are misleading because they lack absolute volume figures and fail to represent the majority of frontier traffic. In terms of inference spending on the platform, the combined costs for Moonshot AI, DeepSeek, and Z.ai surpass those of OpenAI, the second largest spend category.