Open Brain
VercelEngineering PostOfficial Source

Decision models take 26% of requests; Anthropic now wins new teams 2:1

Vercel’s October gateway data shows decision models taking 26% of requests but 0.53% of spend, while model use shifts quickly. Treat routing and model choice as workloads to revisit, not fixed defaults.

Vercel · Nov 9, 2026
Open Source Open MarkdownOpen JSON
Source Summary

In Vercel’s October 2026 gateway data, decision models handled **26% of requests** and **0.53% of spend**; they classify, score, or choose rather than generate text. About 11% of paid teams used voice models, and more than half of language-model requests went to models under 60 days old.

Practical Implication

For agent builders, consider routing simple classification or selection steps to decision models, and budget time to reassess model choices as releases change. Vercel reports that teams processing at least 10 million tokens a month often changed their most-used model between September and October.

Agent-Ready Context
In Vercel’s October 2026 gateway data, decision models handled **26% of requests** and **0.53% of spend**; they classify, score, or choose rather than generate text. About 11% of paid teams used voice models, and more than half of language-model requests went to models under 60 days old.

For agent builders, consider routing simple classification or selection steps to decision models, and budget time to reassess model choices as releases change. Vercel reports that teams processing at least 10 million tokens a month often changed their most-used model between September and October.

The report covers anonymized traffic routed through Vercel AI Gateway, not the whole market. Spend is estimated at list prices, and the vendor’s model classifications and customer mix shape the results.
Connected Context · Feed7 Judgment

This gateway snapshot adds a large request share but small spend share for decision models, alongside evidence of rapid model turnover among heavy-usage teams. It supports separating selection and classification workloads from generation and revisiting routing choices, while remaining a vendor-specific traffic sample rather than a market-wide account or quality comparison.

Open-weight models take 56% of token volume, Astra doubles Fable 5.1 spendBoth are gateway snapshots; the September report describes open-weight volume share, while this one adds decision-model request and spend shares and signals continued mix changes.DeepSeek overtakes Google on volume, cost per token falls 13.6%Both support workload routing as a factor in gateway usage and spend; the November report adds a distinct request-heavy, low-spend category of decision models.Virtual models now available on AI GatewayThe reported model churn makes centrally managed routing settings relevant; Virtual Models provide a task-ID mechanism for those settings, though the report does not establish that surveyed teams used it.CFOs and the new economics of AIBoth describe teams using multiple models, but the Vercel report adds a specific observation that heavy-usage teams often changed their most-used model month to month.
Context Map
industrydata#model-selection#adoption
Uncertainty
The report covers anonymized traffic routed through Vercel AI Gateway, not the whole market. Spend is estimated at list prices, and the vendor’s model classifications and customer mix shape the results.