Two months ago, Vercel’s AI Gateway data showed DeepSeek carrying 17% of gateway token volume while collecting roughly 1% of the spend. The June numbers, published in Vercel’s July index, show that gap has not narrowed. It has widened, and it now describes an entire category rather than one lab.
Open-weight models carried 29% of all tokens routed through AI Gateway in June, up from 11% in April, while taking under 4% of the dollars spent. Vercel’s AI Gateway routes production LLM traffic between applications and labs and publishes monthly aggregates of that traffic. Roughly one in eight enterprise customers on the platform now runs an open-weight model in production, according to the report.
DeepSeek, the Chinese lab known for training frontier-comparable models at a fraction of typical cost, did most of the pulling. It climbed to 22.6% of gateway token volume in June, trailing Google by less than two percentage points for second place behind Anthropic. Google’s own share slipped to 24% as an April surge unwound. Vercel’s report states that an open-weight provider could overtake Google for second place in volume if the current trajectory holds.
A newer entrant shows how fast that trajectory can move. Z.ai, the Chinese lab behind the open-weight GLM model family, released GLM 5.2 on June 16, an MIT-licensed model built for long-running agent workloads and priced at roughly one-fifth of Anthropic’s Opus 4.8. Daily token volume climbed roughly 50-fold in two weeks, and the model reached 11th place on the gateway by tokens in its final week, hitting 7th on individual days. It captured over three-quarters of Z.ai’s entire June token volume within two weeks of launch. Vercel’s report notes that Gemini 3.1 Pro, the fastest in-family migration it had previously tracked, took a full second month to reach a comparable share.
None of that has dented Anthropic’s grip on the money. Anthropic collected 61% of gateway spend in June while running just 32% of tokens, and it took 72% or more of spend in every use case the report classifies as high-stakes: coding agents, back-office agents, and app generation. The top four frontier US labs together took 95% of spend. The pattern is routing discipline, not indecision: cheap, high-volume work goes to open-weight and discount models, while the work where an error is expensive stays on Anthropic’s frontier stack regardless of price.
The report is worth reading with one caveat attached. It reflects traffic through one company’s gateway, not the AI market as a whole, and enterprises that route directly to a lab’s own API rather than through Vercel would not show up in these figures. That limits how far the 29% figure should be extrapolated, even as the direction of the trend across three consecutive monthly reports is now hard to dispute.
For any team currently budgeting AI spend by model rather than by task, June’s data is the argument for splitting that budget in two. High-volume, low-stakes generation should already be priced against DeepSeek and GLM 5.2 rather than a frontier default, while the coding-agent and back-office workloads where mistakes are costly should stay parked on Anthropic regardless of the discount available elsewhere.
Figures reported by Vercel in its AI Gateway Production Index for July 2026, covering AI Gateway usage data from June 2026.
