Vercel’s gateway data separates cheap decisions from expensive generation
Vercel’s gateway data separates cheap decisions from expensive generation
Vercel’s November 9 AI Gateway Production Index reports that decision models accounted for 26% of its October requests but just 0.53% of estimated spending. The gap suggests that request counts alone can badly misrepresent the economics of a mixed-model application. Vercel report
Nearly a third of paid gateway teams used a decision model, according to the company. More than half of language-model requests went to versions less than 60 days old. These are measurements of traffic passing through Vercel, not a census of AI use or proof that one provider leads the whole market.
The report estimates spending at published list prices, which may differ from actual bills. Model age starts on the version’s first day on the gateway; switching statistics cover paid teams meeting its token threshold. Prior periods can change with methodology revisions.
Guth analysis: Separate the operational dashboard into requests, tokens, cost and outcomes. A cheap classifier may dominate volume while a less frequently called generator dominates spending. Neither number tells whether the application did the right thing. Frequent model replacement also makes version-pinned evaluation and regression checks a practical requirement, rather than optional housekeeping.
Status: Newly discovered November 9 report covering October, not breaking November 13 news. No independent audit of Vercel’s underlying traffic was performed.
Sources and citations
Revision history
Published .