Intel CEO Lip-Bu Tan said Chinese frontier AI models such as DeepSeek, Kimi and GLM offer performance that is “good enough,” cost less per token and are being adopted “very fast.” Tan also backed open-source AI as the best way to improve efficiency, and said market demand will determine which models suit different uses. Intel sells chips rather than AI models; Unwire Pro said broader use of lower-cost, open-source models could raise demand for AI hardware and noted that Tan’s comments may reflect Intel’s commercial interests. The article says price differences between models remain clear.

October third-party API pricing data cited by Unwire Pro show DeepSeek V4.1 Flash at US$0.15 per million input tokens and US$0.60 per million output tokens off-peak. For one million input and one million output tokens, that is US$0.75, compared with US$12 for OpenAI’s GPT-6.1 Sol and US$24 for Anthropic’s Claude Opus 5.5; their respective input and output rates are US$2 and US$10, and US$4 and US$20. DeepSeek’s peak-hour rates double to US$0.30 input and US$1.20 output. Other Chinese models are not all as cheap: Kimi K3 charges US$10 per million output tokens, matching GPT-6.1 Sol, while GLM 5.3 charges US$4.40, or 44% of GPT-6.1 Sol’s rate.

Yicai reported that Chinese models processed approximately 61.2 trillion tokens on OpenRouter during the week of September 7–13, nearly three times the 21.8 trillion attributed to U.S. models. Chinese models had led U.S. models on the platform for 20 consecutive weeks, with lower costs cited as a factor in developers’ choices. The figures cover only OpenRouter traffic, not the broader AI market. The Washington Post reported that U.S. firms DoorDash and Airbnb use customizable, lower-cost open-weight models from Chinese companies, including Moonshot AI, DeepSeek and Alibaba, for routine tasks; U.S. closed models from Anthropic and OpenAI remain more advanced.

Intel CEO Lip-Bu Tan said Chinese frontier AI models such as DeepSeek, Kimi and GLM offer performance that is “good enough,” cost less per token and are being adopted “very fast.” Tan also backed open-source AI as the best way to improve efficiency, and said market demand will determine which models suit different uses. Intel sells chips rather than AI models; Unwire Pro said broader use of lower-cost, open-source models could raise demand for AI hardware and noted that Tan’s comments may reflect Intel’s commercial interests. The article says price differences between models remain clear.

October third-party API pricing data cited by Unwire Pro show DeepSeek V4.1 Flash at US$0.15 per million input tokens and US$0.60 per million output tokens off-peak. For one million input and one million output tokens, that is US$0.75, compared with US$12 for OpenAI’s GPT-6.1 Sol and US$24 for Anthropic’s Claude Opus 5.5; their respective input and output rates are US$2 and US$10, and US$4 and US$20. DeepSeek’s peak-hour rates double to US$0.30 input and US$1.20 output. Other Chinese models are not all as cheap: Kimi K3 charges US$10 per million output tokens, matching GPT-6.1 Sol, while GLM 5.3 charges US$4.40, or 44% of GPT-6.1 Sol’s rate.

Yicai reported that Chinese models processed approximately 61.2 trillion tokens on OpenRouter during the week of September 7–13, nearly three times the 21.8 trillion attributed to U.S. models. Chinese models had led U.S. models on the platform for 20 consecutive weeks, with lower costs cited as a factor in developers’ choices. The figures cover only OpenRouter traffic, not the broader AI market. The Washington Post reported that U.S. firms DoorDash and Airbnb use customizable, lower-cost open-weight models from Chinese companies, including Moonshot AI, DeepSeek and Alibaba, for routine tasks; U.S. closed models from Anthropic and OpenAI remain more advanced.