The takeaway
Decision models are turning agent fleets into a routing problem: most calls are cheap choices, not reasoning, so the win is moving them off frontier models.
Why it matters for builders
Audit agent traces and count how many calls are classification or routing decisions. Move those steps to a small decision model or distilled classifier and keep frontier models for planning, synthesis and user-facing output. The savings compound at volume and latency falls with them.
Decision Models Take Off as Jev Tops Vercel Adoption Data
Jev, the lightweight "decision model" from TypeSafe AI, has become the fastest-adopted model ever measured on Vercel's AI gateway, according to usage data Vercel published on Thursday. Within two weeks of its limited release, 35% of Vercel's paid teams were calling it, nearly double the reach of the second most popular model, Sonnet 5.5.
What the numbers show
Vercel's gateway is a unified API service comparable to OpenRouter, so its data is a partial view - it will undercount labs like OpenAI and Anthropic that serve their own APIs. Even so, it is the first hard usage evidence behind the buzz that has surrounded Jev since its September 15 launch.
The model's trick is narrowness. Jev is built to choose between a few predetermined outputs, which makes it cheaper and faster than general-purpose models and well suited to classifying large volumes of data quickly. Vercel itself replaced GPT-6 Luna with Jev for classifying harmful commands.
The same report shows that open-weight models now handle more than 70% of the tokens processed through Vercel's gateway and account for nearly 25% of total spend. Spending on frontier labs grew tenfold over the same period, which suggests the two tiers are complements rather than substitutes: expensive models for the hard work, small ones for the volume.
Why decision models are everywhere now
Jev's popularity has already pulled competitors in. OpenAI and Amazon have both shipped comparable offerings, and OpenAI's Decisions API is in public beta as a faster alternative to its Responses API for real-time workloads.
That is a real shift for anyone building automation. A large share of agent calls are not reasoning problems at all - they are routing questions: which queue, which label, which tool, which branch. Those calls do not need a frontier model, and paying frontier prices for them is the most common source of waste in production agent fleets, a problem we covered in our analysis of agent crystallization.
What to do with it
Audit your agent traces and count how many calls are classification or routing decisions. If the answer is "most of them," move those steps onto a small decision model or a distilled classifier and keep the frontier model for planning, synthesis, and anything user-facing. Savings compound at volume, and latency drops with them.
For now, Vercel's adoption numbers are the clearest evidence yet that the economics of agent fleets get decided one cheap call at a time.
Sources
The Automation Brief
Read 5 AI stories instead of 50.
The essential moves in AI agents, models, automation and infrastructure — filtered for builders and operators, with the part that actually matters.
No noise. Unsubscribe anytime.
Editorial notes
Stefan Trbojevic
n8n Lab Editorial
16 October 2026
16 October 2026
Sources
AI disclosure: AI assisted with research and drafting. Factual claims are reviewed by an editor.
