The cost for businesses to operate AI models has fallen to its lowest point in 2026, according to research by investment bank Jefferies. This decline is attributed to a heated global price war and increased use of low-cost Chinese open-source AI tools, including those from DeepSeek.

From August 6 to 8, average inference prices—measured per million tokens processed by a model—ranged between US$1.16 and US$1.18. Jefferies cited data from US research firm Silicon Data, noting this as the lowest level recorded this year. The average unit costs have dropped sharply from US$2.04 on May 31 and US$1.45 in late July.

Silicon Data’s index monitors pricing across business application programming interface (API) providers and open-weight inference platforms used by software developers. Jefferies analysts, led by Thomas Chong, highlighted that this price decline aligns with an increasing focus on cost efficiencies in both US and Chinese tech ecosystems.

OpenAI intensified the price competition last month by cutting rates for its GPT-5.6 model series by up to 80%. Similarly, Anthropic’s Claude Opus 5 offers performance comparable to its flagship Fable 5 model at half the price, according to a Jefferies report released last month.

Chinese firms continue to push the boundaries of affordable computing on the open-source front, contributing to the overall reduction in enterprise AI costs.

Sources