DeepSeek overtakes Google on volume, cost per token falls 13.6%
DeepSeek just became the second-largest lab by token volume. Cost per token fell 13.6% in July alone. Here's what actually happened in enterprise AI.
Why it matters
Market-shifting data: Open-weight models (DeepSeek, Moonshot, Z.ai) are capturing production volume at commodity pricing, forcing frontier labs into a two-tier market. Anthropic holds 65% of spend on 30% of volume by pricing power alone; cheaper alternatives exist but require task-switching. This is the first month where open-weight models meaningfully competed on capability-per-dollar, not just cost.
The key facts
11 to knowDeepSeek captured 25% of gateway token volume in July, surpassing Google's 11% (down from Google's 40% in April)
DeepSeek V4 Flash ran 19% of all tokens on the gateway, 70% more than the next model
Open-weight models' share of gateway spend more than doubled to 8.6% in July, driven by Kimi K3 and GLM 5.2
Average price per token fell 13.6% in July despite 37% spend growth, driven by model mix shift not price cuts
Anthropic collected 65.1% of gateway spend on 30% of token volume at 4.4x average price per token
Kimi K3 (released July 16) scaled to 8th place by volume within two weeks, tripled daily volume by month end
81% of July tokens ran on models not present on gateway six months prior
Three in four teams with >10M monthly tokens changed their model mix by at least 10%; half changed by 25%+
Anthropic's Haiku 4.5 (cheapest offering) costs 68% of gateway average; DeepSeek V4 Flash costs 6% of that average
In coding agents (largest workload): DeepSeek ran 32% of volume, Anthropic collected 80%+ of spend
Moonshot's share of total gateway spend quadrupled to 2.3% in July
Go to the source
Vercel Blogvercel.com
Publisher excerpt: AI Gateway Production Index — August 2026 Every month, routes tens of trillions of tokens between production applications and AI labs. That traffic gives us a view of what AI usage actually looks like in today's enterprise, and we publish it here monthly. See the Production Index reports from , ,…