Open-weight models hit 56% of AI Gateway tokens

Vercel's AI Gateway Production Index tracks anonymized traffic routed through its gateway, which the company says routes tens of trillions of tokens monthly between production applications and AI labs. The September 2026 report covers data collected through August 2026, and includes updates logged between editions on September 17 and September 18. The core shift it documents is a pricing and workload migration: open-weight models, which processed fewer than one in ten tokens in December 2025, rose every month from April through August to reach 56% of volume. Over the same period, the average price per token fell 23.2% in August, the third straight monthly drop, and the median team running more than ten million tokens paid 7.6% less.

The report also traces how spending moved within Anthropic's lineup after Opus 5 came online at the end of July. Fable 5's share of gateway spend had surged to 13.2% in July when its access was restored under a lifted US export control, but fell to 4.9% in August as Opus 5's share rose to 22.5%. Nine in ten teams that ran Fable cut their usage, and Anthropic still kept 64 cents of every dollar spent in August because those workloads stepped down to Opus 5 rather than to another lab. A similar pattern played out at Google, where Gemini 3 Flash lost 95% of its share of gateway tokens since May and more than three-quarters of the volume it lost went to other labs.

The stakes here hinge on whether the migration toward open-weight and cheaper-tier models continues into September and beyond, and on whether frontier labs can keep premium spend even as their most expensive models lose share. For buyers running production AI workloads, the practical question is whether the gap between open-weight and frontier capability keeps narrowing enough to justify moving more tasks off the premium tiers.