Open-weight models carried over 70% of September's tokens on Vercel's AI Gateway. They took about 23% of the dollars.

That gap is the story. Vercel's October index shows a barbell, not a winner: one cheap model took nearly half of all tokens for 2.6% of spend, while another open model took 3.6% of tokens and 13.2% of spend.

Why it matters if you build on LLM APIs: your bill is decided by which model each request goes to, not by which category of model you picked.

What the data says, in three lines:
- Teams past 10M tokens used 15 models on average (Vercel)
- 68% of teams raising open-weight spend also raised closed-weight spend
- Nearly half of teams changed their most-used model in a single month

The layer that matters is the router, plus an acceptance check that tells you what an accepted result costs. Vercel's index can't show that part, so I drew it in.

Caveats: one gateway, list-price spend estimates, and a couple of internal inconsistencies in the report.

Question for you: if you route between models, what decides the tier: task type, risk, or something else?

Full breakdown with diagram: [link to Medium article]

#AIEngineering #LLM #MLOps