Decision models go from nothing to 26% of Vercel gateway traffic

Vercel published its AI Gateway Production Index on 9 November, built from anonymised traffic across a gateway it says now carries over 100 trillion tokens a month. The headline number: decision models, the small classification and routing models used to pick an action rather than hold a conversation, went from zero requests in August to 26% of all gateway requests by October. They cost roughly a seventieth as much per request as a language model call, so despite the volume they’re still only 0.53% of total spend.

The report also found more than half of October’s requests went to models under 60 days old, up from under 30% previously, and that three in four teams who try a new model have moved off it again within a month of its release. On provider share, Anthropic and OpenAI are now level on team reach, at 55 and 54 teams per 100 respectively, and new teams picked Anthropic nearly two to one in October, even as Anthropic’s share of overall spend roughly halved since December while OpenAI’s rose.

Why it matters: if most teams are off a model within a month, pinning your stack to one model’s quirks is a short-lived bet. Cheap decision models are now a mainstream way to cut routing costs without touching your main LLM spend.

The caveat: this is Vercel’s own gateway traffic, not the whole market, so it best reflects teams already routing through Vercel rather than every developer using these providers directly.