DeepSeek leads surge in low-cost Chinese open-weight models on US platform

DeepSeek's newest light-weight AI model has led a surge in the usage of open-weight Chinese models on the US web development platform Vercel, according to data released by the company. Open-weight models now make up 54% of token volume on Vercel's AI Gateway, surpassing proprietary models which account for 46%. This marks a significant shift from June 24 when open-weight models only accounted for 28% of the token volume, compared to 72% for proprietary models.

DeepSeek-V4-Flash is the most-used model according to token volume, while Chinese models including Step 3.7 Flash, GLM-5.2, and Z.ai dominate the top five. OpenAI's GPT-5.6 Luna also features in the top five. This trend suggests developers are favoring cheaper and open models for production workloads, especially for autonomous agents that require large amounts of tokens for tasks like reasoning, coding, and tool usage.

This democratization of AI, driven by open-source models, is also a concern for security, as hacking groups have started using models like DeepSeek for automating tasks and developing malware. While the data reflects activity on Vercel's platform, separate corporate spending data shows businesses are becoming more price-sensitive towards advanced proprietary models.

Anthropic's Fable 5, their most expensive model, only generated 11.4% of spending in the month after its release, compared to 25% for OpenAI's GPT-5.6 Sol. This shift is also prompting US investment in domestic open-weight models, with Nvidia planning to use a $6 billion licensing agreement with Poolside to build a powerful open-weight model that could challenge DeepSeek and Moonshot AI's Kimi K3.