Jev, a model that only picks from fixed answers, reached 35% of Vercel's paid teams in two weeks. Use it for the yes or no steps

A lot of the steps in a business automation are not writing tasks. They are yes or no questions.

What happened

TechCrunch reports that Jev, a decision model from TypeSafe AI, was used by 35 percent of paid teams on Vercel's AI gateway within two weeks of its launch, nearly double the reach of the second most popular model, Sonnet 5.5. Vercel's data calls it the fastest two week adoption it has seen. Jev went into limited release on September 15. It is built to choose between a few predetermined outputs, which lets it run cheaper and faster than general models, and Vercel itself uses it in place of GPT-6 Luna to classify harmful commands. TechCrunch notes the gateway view may undercount labs that people use through their own APIs.

LangChain's guide to agent evals adds a cost data point. In its Jev-as-a-Judge experiment, Jev matched human labels on every pass or fail decision and cost $0.34 in total, compared with $28.17 for Claude Sonnet 4.6. The same guide recommends binary pass or fail or a small set of labels over 1 to 10 scales, because they are far more reproducible.

My take

Look at any client workflow and count the steps that are really classification. Is this lead a fit. Is this email a complaint. Does this ticket need a human. Did the reply include the price.

Those steps are often running on a large general model because that is what was wired in first. That is paying writing prices for sorting work.

How I would approach it:

1. List the decision steps in one workflow and give each a fixed set of answers.
2. Run a cheap decision model on them, and keep the big model for steps that write.
3. Check it against 20 or so real examples a person has already labeled before switching.
4. Log every decision so you can spot drift.

One caution: that test result is from LangChain's own experiment. Run your own small comparison before you trust any model with a routing step.