Chinese-built AI models have captured over 30% of token usage among U.S. companies weekly since February 8, with usage peaking at 46%, according to OpenRouter data. This marks a dramatic increase from 11% over the previous 12 months.
Chinese AI models are rapidly gaining adoption among American companies, reflecting narrowing performance gaps with domestic alternatives.
OpenRouter, a platform that routes requests across multiple AI models, tracks token consumption patterns across its user base. The data reveals a sharp acceleration in Chinese model usage beginning February 8, with weekly token share consistently exceeding 30% since that date.
The peak of 46% represents a four-fold increase from the 11% baseline established over the preceding year. This trajectory suggests either rapid performance improvements in Chinese models, cost advantages driving adoption, or both factors combining.
The shift carries implications for U.S. technology strategy and AI development priorities. Chinese models gaining significant traction in American enterprises indicates the competitive landscape has shifted materially in recent months.
Key Chinese AI models competing in this space include Alibaba's Qwen, Baidu's Ernie, and others that have made substantial improvements in reasoning, coding, and multimodal capabilities. Several now rival or exceed some American alternatives in benchmark performance.
Cost differentials may also play a role. Chinese providers often price API access below U.S. competitors, making them attractive for cost-sensitive applications or high-volume token consumption.
The timing coincides with increased focus on AI capabilities globally and potential shifts in regulatory or trade policy affecting technology procurement. U.S. companies may be diversifying model providers or testing alternatives amid evolving geopolitical considerations.
OpenRouter's data covers its platform only, which represents a subset of overall AI usage. Enterprise deployments using direct API access to OpenAI, Anthropic, or other providers would not appear in these figures. However, the platform captures significant usage patterns among developers and companies evaluating multiple models.
The data suggests the AI market is becoming less concentrated, with Chinese models transitioning from niche to mainstream adoption among American businesses seeking cost-effective or specialized capabilities.
Former Google China chief Kai-Fu Lee says Chinese AI rivals pose a serious threat to OpenAI and Anthropic, citing lower costs and aggressive business models. The assessment highlights growing competition in the global AI market.
OpenAI's GPT-6 Astra achieved human-level efficiency on the ARC-AGI-3 benchmark for the first time, prompting ARC Prize chief François Chollet to accelerate his AGI forecast. However, benchmark disagreement clouds the broader picture of the model's capabilities.
Instagram's AI content labels are malfunctioning, incorrectly flagging user photos as AI-generated while missing actual synthetic imagery. The reliability issues undermine Meta's effort to combat misinformation on the platform.
Claude can reference previous conversations to inform current interactions. Keeping that conversation history accurate is critical for reliable AI assistance.