Average prices for AI inference fell to their lowest level this year, ranging from $1.16 (£0.86) to $1.18 per million tokens from 6 to 8 August, according to investment bank Jeffries, citing analysis from Silicon Data. This is significantly below the $2.04 average on 31 May and the $1.45 average of late July. The drop was driven by the release of low-cost open-weight Chinese models and an increased focus on price-efficiency by US developers. The latest decline follows the release of V4-Flash-0731 from Hangzhou-based DeepSeek, which costs $0.03 per task and was described as “materially lower” than competitors. The model rose to the top of global token consumption rankings at OpenRouter, with 27 percent of total processing volume on the platform as of Monday, compared to Google’s 25 percent. DeepSeek’s model handled 8.22 trillion tokens last week, followed by Tencent’s Hy3 at 7.13 trillion and a version of V4-Flash released in April at 6.05 trillion. Jeffries analysts said the downward pricing trend coincided with an “increasing emphasis on cost efficiencies” in both the US and China. Last month OpenAI slashed prices for its GPT-5.6 model series by up to 80 percent, while Anthropic’s Claude Opus 5 delivered comparable performance at half the price. Despite