A million words costs $50 from Anthropic and 87 cents from DeepSeek. Corporate America noticed.
Corporate U.S. firms are shifting to cheaper AI models, especially Chinese ones, to cut token costs while balancing performance.Summary
Roughly 750,000 words of AI output cost $50 from Anthropic’s Fable but only 87 cents from DeepSeek’s V4‑Pro, prompting U.S. companies to re‑evaluate their model choices. Firms that once competed to maximize token usage now prioritize thrift‑maxxing, capping staff spend and substituting cheaper models such as Z.AI’s GLM‑5.2 or Moonshot’s Kimi K3, which offer comparable performance for a fraction of the price. This shift has led to a broader adoption of Chinese open‑weight models, with 57% of tokens consumed by U.S. firms on OpenRouter in a single week coming from Chinese models and 47% of surveyed decision‑makers reporting at least one use case.
The price gap is driven by lower electricity costs, less regulatory resistance, and Chinese firms’ willingness to operate on thin margins. Export controls have also forced Chinese labs to innovate on less‑advanced hardware, further reducing costs. As OpenAI and Anthropic prepare for public listings, the surge in demand for inexpensive models threatens the valuations built on high‑capability offerings and fuels lobbying battles over Chinese open‑weight models. While some executives view the trend as a normal market adjustment, concerns remain about reputational risk, procurement pushback, and the potential impact on the U.S. AI race. The upcoming release of Moonshot’s K3 weights and the continued rise in token usage suggest that the market will keep refining its model mix to balance cost, performance, and geopolitical considerations.
(Source：Tnw)