Open models hit 78.4% of tokens on Vercel AI Gateway

Guillermo Rauch posted what he called a possible record day for open weight models on Vercel AI Gateway, with open models at 78.4% of token volume against 21.6% closed. Spend tells a different story, as it usually does, but not by as much as it once did. Third and fourth by spend that day were Moonshot AI and DeepSeek, and Rauch says adding Z.ai puts their combined spend past OpenAI, which sat at number two.

One caveat he flags himself. That is spend on inference across providers, mostly in the US, not revenue going to the open weight labs.

The routing habit is spreading

The gateway numbers line up with what individual engineers say they are doing by hand. Matteo Collina says 80% of the code he produces comes out of OpenAI Luna or DeepSeek Flash, while 80% of the value he generates comes from Fable or Sol. His split is not about price alone. If he knows what he is doing, he says the fast and cheap models are the better choice, and he saves the expensive ones for when he needs a design partner.

Matteo Collina

@matteocollina

X

if I know what I’m doing, the fast&cheap models are better choice. If I need a design partner, I go for the best I can afford.

David Cramer goes further and suggests people are paying for a difference they cannot measure. He argues that anyone on Fable or Astra should try switching back to Opus or Sol and see whether their tasks perform any differently, and predicts most will find they do not.

That is an assertion rather than a benchmark, and Cramer does not show numbers. Collina's split is one engineer's own workflow. But it is the same shape as the budget story we covered earlier, where Gergely Orosz described an AI-pilled company that had unlimited AI budgets a year ago and now runs a daily budget limit on Astra and Fable, with Sol and Opus still unlimited. Their instruction to engineers was to use the two capped models only for planning.

For working engineers, none of this says the frontier models got worse. It says the gap between the top tier and the tier below it is no longer wide enough to justify paying frontier prices for routine churn, and that the people who watch token volume at gateway scale are seeing the same thing as the people watching their own invoices. Whether the cheap models genuinely match the expensive ones on hard work is still argued from anecdote on both sides.

David Cramer

@zeeg

X

you'll likely realize your tasks dont perform any differently