AI
Open-weight models surge to majority of tokens, but Anthropic holds 64% of spending
Despite handling more than half of token volume on Vercel's AI Gateway, cheaper open-weight models account for only 14 cents per dollar of actual spending.
The New Stack
Open-weight models have crossed a major threshold: they now process the majority of inference requests flowing through Vercel's AI Gateway, yet the economics tell a starkly different story about which models are actually driving spending.
In August, open-weight models like those from DeepSeek and Moonshot AI processed 56% of all tokens routed through Vercel's gateway—a watershed moment that marks their first majority share in monthly volume. Just eight months earlier, in December 2025, these models accounted for barely 7% of traffic. The climb has been steep and consistent: the share reached 13% by April, then accelerated every month after, hitting 36% in July before crossing the 50% threshold in August.
The token-to-dollar disconnect#
The headline numbers mask a crucial economic reality. While open-weight models handled more than half of computational work, they generated only 14 cents of every dollar spent through the gateway—a fundamental disconnect between usage and revenue.
This gap reflects the underlying pricing dynamics of the AI market. Open-weight models are substantially cheaper to operate than proprietary offerings from US frontier labs. Vercel's data shows that across its platform, average token prices fell 23.2% in August alone, with this being the third consecutive month of price declines.
Anthropic's fortress at 64%#
Anthropic maintains a dominant position in spending, claiming 64 cents of every gateway dollar in August. The Claude maker has shown remarkable consistency since December 2025, never dropping below 61% of spending and consistently occupying the top revenue positions.
Notably, Anthropic's dominance persists even as customers migrate between its own models. Claude Fable 5 declined sharply from 13.2% to 4.9% of gateway spending between July and August, while the cheaper Claude Opus 5 climbed to 22.5%. Rather than losing revenue, Anthropic simply captured it through different tiers of the same family. Opus 5 ultimately gained roughly twice the volume that Fable 5 lost—a testament to Anthropic's ability to retain spending even when customers downshift to lower-cost alternatives.
Switching is real, but loyalty follows capability#
Vercel's data reveals that customers will readily cross brand boundaries when a new model fails to deliver. More than three-quarters of the volume Google lost from Gemini 3 Flash migrated to competitors including OpenAI and Anthropic. Google's overall token share on the gateway collapsed from 30% to just 5%—with the Gemini 3 Flash decline alone accounting for 22 of those 25 percentage points.

