Underclassic / explainer

Open-weight models: 56% of tokens, 14% of spend

Published 2026-10-12. Sources: Vercel AI Gateway Production Index, September 2026; Ramp AI Index, September 2026.

56%
of tokens on Vercel AI Gateway ran on open-weight models in August 2026

14%
of gateway spend went to those same open-weight models

6.4%
of AI-spending U.S. businesses on Ramp use open-model routing

What the numbers say

Vercel routes production AI traffic for thousands of teams through its AI Gateway. In its September 2026 index (data through August), open-weight models ran 56% of all gateway tokens, the first month they took a majority. In December 2025 they handled 7%; from April to August their share rose every month, from 13% to 56%.

The money did not follow at the same rate. Vercel says open-weight models processed 56% of August's tokens and accounted for 14% of spend. Anthropic alone took 64 cents of every dollar spent through the gateway in August.

The cheap volume is pulling prices down: the average price per token on the gateway fell 23.2% in August, the third straight monthly drop, and the median team (above ten million tokens a month) paid 7.6% less per token.

Why tokens and dollars split

Open-weight models are priced like commodities. On our model board, open rows such as DeepSeek V4.1 Flash and GLM-5.3-Flash list at well under a dollar per million input tokens, while frontier closed models list at $2 to $10. So a model family can carry most of the work and still be a small slice of the bill. Teams send the bulk, repetitive jobs to cheap open models and keep the frontier tier for tasks that justify the premium, which is roughly how Vercel describes it.

One example from the same report: within five days of Z.ai launching GLM-5.3-Flash, it was running three times GLM-5.2's daily volume, and by August 31 it processed two-thirds of Z.ai's tokens on the gateway.

The other side: most businesses haven't switched

Gateway data skews toward developer teams. Ramp's card-spend data looks at businesses broadly, and there the open-model shift is small. In its September 2026 AI Index, Ramp says only 6.4% of AI-spending businesses use open models, and 3.6% of businesses overall. Ramp measures this through routing platforms that also sell closed models, so it says actual open-source adoption is likely lower. Ramp also tracks an effective price per million tokens that has fallen 41% to $0.68 from a 2026 peak of $1.15 in March.

What it means

Usage is going open faster than revenue is. For the closed labs, that is a volume problem hidden by a spend lead: they keep most of the money while losing the majority of the work. For anyone running AI themselves, it is the case for local and self-hosted inference: the workloads most teams are moving to open weights are the ones you can run on your own hardware. See our buyer's guide to home AI computers.

Caveats

Vercel's figures cover traffic on its own gateway only, and spend is estimated from labs' published list prices; actual bills may differ.
Vercel's token volume includes input, output, reasoning and cached tokens, and its open-weight classification is broader than in its earlier reports.
Ramp's 6.4% is measured by adoption of routing platforms, which Ramp says likely overstates true open-model use.

Subscribe

$ echo you@email.com | underclassic --subscribe

One email per issue. No tracking pixels, no filler.

Read today's issue · Model price board