FoxTPNL / Wikimedia Commons (CC BY 4.0) Open-weight models surged from 11% to 56% of token volume in just four months, but closed models still capture the lion's share of revenue
The AI model landscape just hit a tipping point. Open-weight models now account for 56% of all token volume flowing through Vercel’s AI Gateway as of August 2026, up from a mere 11% in April. That’s a fivefold jump in four months, and it marks the first time open models have overtaken their closed counterparts on this metric.
On a single day, August 22, the share spiked to roughly 62%. More recent leaderboard data has shown it touching 78.4%.
The volume-revenue paradox #
Before anyone declares total victory for the open-source crowd, there’s a wrinkle worth examining. Anthropic, maker of the Claude family of closed models, still commands 61-65% of total gateway spend despite representing only about 30% of tokens processed. Its models carry premiums of up to 4.4 times the average token price.
That’s the “barbell” pattern emerging across enterprise AI usage. On one end: massive volumes of cheap inference running through open-weight models from labs like DeepSeek, Z.ai, and Moonshot. On the other: premium, high-stakes workloads where companies pay a significant markup for Anthropic’s quality guarantees.
The news moving money, markets, and the world—before your day starts.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
DeepSeek has become particularly prominent in this shift, frequently ranking first or second in volume contribution with roughly 25% or more of total token share. Google, by comparison, sits at around 11%.
A race to the bottom on price #
The economics driving this shift are stark. Average price per token fell 23.2% in August alone, the third consecutive monthly decline. Over the past five months, prices have dropped more than 50%.
Open-weight models are running high-volume workloads at approximately one-seventh the cost of frontier closed models.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our