DeepSeek Surpasses Google in Usage; Cost per Token Drops 13.6%

Vercel AI Gateway released its August Production Index based on July 2026 data. As token consumption passing through the gateway grew rapidly, total spending increased by 37%, but the average price paid per token dropped by 13.6% compared to the previous month. Prices rose by about 20% in May and remained stable in June.

DeepSeek surpassed Google in token throughput, ranking 2nd among all AI labs. DeepSeek's token volume was recorded as more than twice that of Google.

In terms of cost, Anthropic accounted for 65% of gateway spending, but only 30% of the token share. The average price of Anthropic tokens was 4.4 times higher than the average of other labs.

The rankings for media models also shifted.

• Google's Nano Banana overtook OpenAI's GPT Image in image throughput.
• ByteDance's Seedance took first place in both video throughput and revenue.

Kimi K3, released by Moonshot on July 16, is a model targeting long-running agent tasks, using approximately 12 times more tokens per request than its predecessor, K2.5. Daily throughput tripled from the launch week to the last week of July, and on the last day of July, it ranked 8th in gateway token volume.

Within two weeks, K3 accounted for nearly two-thirds of all Kimi tokens, expanding its share to 82% in the last week of July. The share of gateway spending for open-weight models also more than doubled to 8.6% in July, with Moonshot and Z.ai accounting for over 90% of this increase.