Jev: a model that only picks from a list

Jev is a new model from TypeSafe. It reads text like an LLM, but it can't write anything back. You give it the possible answers up front and it picks one. That sounds limiting, yet within 24 hours of launching on September 15 it was the fastest-adopted model in the history of Vercel's AI Gateway.

A missing building block

Software has had deterministic code for decades: if an invoice is 30 days overdue, flag it. LLMs added reasoning and writing. What was missing was a fast, cheap, general-purpose classifier: something that understands messy text and returns a simple choice.

A huge number of everyday problems have that shape. Does this customer sound like they're about to cancel? Is this email a real opportunity or just using the words? Should this agent be allowed to run that command? Until now the options were custom machine-learning classifiers (data labelling, training and maintenance for every new category) or an expensive LLM call to make a multiple-choice decision.

Jev covers that gap. You describe the question and the allowed answers, and it returns a choice with scores. It works across very different problems without retraining, and it can answer several questions in one call.

Four ways it fits into software

• Between messy input and existing code. A support ticket comes in; Jev tags it as billing, urgent and a churn risk, and your normal code routes it. The same goes for emails, or for flagging a risky agent action such as a force push before it runs.
• Finding what matters in a large pile. One scientist had Jev pick the 100 most important immunology questions out of 10,000 candidates. Jev doesn't answer them; it helps decide where to spend attention.
• Choosing the next step. Jev runs the outer loop of a workflow and decides whether to call a tool, a cheap model, a frontier model or a human. Browser agents can work the same way, choosing which link or button to use next.
• Intelligence inside ordinary features. In one demo, typing "Urgency" as a spreadsheet column header made Jev classify every row instantly. Add columns for missing information or the right team, and normal formulas can combine those judgments with dates and amounts.

The numbers

• One developer moved an existing tax-document pipeline from an LLM to Jev and reported it was 34 times cheaper and 6 times faster.
• Another sorted 20,000 of his own emails, Slack messages and transcripts in 7 minutes for about $1.
• Published pricing is 4.2 cents per million input tokens, with nothing charged for output. At 1,000 input tokens per request, a million requests cost about $42.
• TypeSafe's launch evaluation reports roughly 100 times faster and more than 100 times cheaper than an LLM for the same decisions.

It doesn't replace LLMs

Jev still makes mistakes, so test it on your own problems against the LLM call or classifier it would replace. Anything that needs reasoning, writing or explaining is still LLM work. Cheap classification may even create more of that work, because it surfaces more exceptions and opportunities worth a closer look.

The name comes from Jevons paradox: when a resource gets much cheaper, we end up using far more of it. Checks that were too expensive to run once per document can now run on every section, every customer and every step of an agent's task.

How to try it

1. Copy the setup prompt from TypeSafe's site into Claude Code, Codex or another agent, create an account and generate an API key.
2. Ask the agent to find a place in your project where an LLM is choosing between fixed outcomes, and build a Jev version.
3. Compare accuracy, speed and cost between the two.

A useful way to spot candidates: anywhere software reads something complicated and only needs to pick from a few outcomes. Look for decisions hidden inside expensive AI calls today, and for judgments you skipped entirely because they used to cost too much.