AI Tools · 6 min read
What Is Jev? The Essay-vs-MCQ Explanation, Plus the Free Setup
ChatGPT writes essays. Jev only does multiple choice — that one sentence explains the most hyped AI model of the month. Here is the plain-language explanation, and the exact three-step setup to use it free on Vercel's AI Gateway.
In a hurry?
Jev is free on Vercel's AI Gateway until September 25 (after that: $0.042 per million input tokens, output free). If you just want to run it, jump straight to the setup.
Skip to the free setup ↓"Key Takeaways"
- Jev is a 'System One' model from TypeSafe AI ($40M seed, founded by an ex-OpenAI researcher). It does not generate text — it returns typed decisions: yes/no, a pick from options, or a score.
- The simplest frame: ChatGPT and Claude write essay answers, word by word. Jev only does multiple choice — it picks from options you define.
- Every answer carries a calibrated confidence score, so your app knows when to act automatically and when to route to a human.
- It answers in as little as 70 milliseconds and can't hallucinate text, because it can't write text. It can still be wrong — that is what the confidence score is for.
- Setup is three steps on Vercel's AI Gateway; the exact code is at the bottom of this page.
Summarize this article with
The one-sentence version
Everyone is arguing about Jev; almost nobody explains it simply. So here is the board-exam version: ChatGPT writes essay answers. Jev only does multiple choice.
If that sentence clicked, the rest of this page is just detail — including the part most coverage skips, and the exact code to try it free.
How LLMs answer: essays, one word at a time
ChatGPT, Claude, Gemini — every chat model generates its answer word by word, each word predicted from the ones before it. That is what makes them expressive, and it is also the root of their three production headaches: it is slow (seconds, not milliseconds), it is priced per generated token, and occasionally the essay drifts into fiction — hallucination.
For writing, that trade is fine. But most of what companies actually call AI APIs for in production is not writing. It is deciding: route this ticket to billing or support? Score this lead 1 to 10? Is this comment spam? Paying an essay-writer to answer multiple-choice questions is the most expensive way to do it.
How Jev answers: multiple choice only
Jev refuses to write essays. You hand it your app's state — a support transcript, an order object, a comment — plus questions with predefined answer shapes: a yes/no, a choice from up to 255 options, or a score on a rubric. It evaluates all your questions in one parallel pass and returns a typed answer for each.
No generation means no waiting: answers land in as little as 70 milliseconds. And because the output is constrained to your options, there is nothing to parse and no essay to drift into fiction. It cannot hallucinate text — it literally cannot produce text.
The underrated part: confidence scores
Honest caveat first: a model that can't write can still pick the wrong option. Jev's answer to that is calibration — every response carries a probability that reflects how sure it is (trained with a method TypeSafe calls RLCD, reinforcement learning for calibrated decisions).
That enables the pattern that makes AI safe to put in production: above roughly 0.9 confidence, automate. Between 0.5 and 0.9, route to a human. Your app finally knows when to trust the AI — which is more than you can say for an essay that always sounds confident.
What it can and can't do
- Can't: write emails, code, summaries, or any text. Essays stay with ChatGPT and Claude — Jev replaces none of that.
- Can: route tickets, score leads, flag content, verify another LLM's output, power real-time UI logic — the millions of small decisions your systems make daily.
- Cost after the free window: $0.042 per million input tokens, and output tokens are free. Context: 64k tokens per request.
- Speed: 70–500ms end to end; multiple questions answered in one parallel pass.
Use Jev free: the 3-step setup
- Prereqs: a free Vercel account (vercel.com) and Node 22+. Jev is free on Vercel's AI Gateway until September 25; after that the pricing above applies — same setup either way.
- In your Next.js project: pnpm i ai (that's AI SDK 7.0.105+). Then run: vercel link, then: vercel env pull — this creates the VERCEL_OIDC_TOKEN that authenticates your calls.
- Call evaluate() with the model id typesafe-ai/jev and your typed questions — copy the exact code below.
The exact code
Paste this into your project. It asks one boolean question about a support transcript; swap the state and question for your own use case:
import { experimental_evaluate as evaluate } from 'ai'; export async function wasRefunded(transcript) { const result = await evaluate({ model: 'typesafe-ai/jev', state: transcript, questions: { refunded: { type: 'boolean', instructions: 'Was a refund issued to the customer?', criteria: { true: 'The agent confirmed money was returned to the customer.', false: 'No refund was issued, or the refund was declined.', }, }, }, }); return result.refunded; // { probability: 0.97 } -> strong yes }
The two gotchas
Local auth tokens expire every 12 hours. If a request suddenly returns 401, just re-run: vercel env pull. Deployments on Vercel refresh the token automatically.
Question types are boolean (yes/no with probability), choice (up to 255 options), and score (2–10 levels). Define options carefully — Jev can only be as good as the choices you give it.
Quick questions
Is Jev a ChatGPT replacement?
No. It cannot write anything. It replaces the decision-shaped calls you may currently be sending to chat models — routing, scoring, classification — not the writing ones.
If it can't hallucinate, is it always right?
No — it can pick the wrong option. What it can't do is invent text. The calibrated confidence score exists precisely so your app knows when an answer is trustworthy and when it needs human review.
What happens after September 25?
The free window on Vercel's AI Gateway ends and standard pricing applies: $0.042 per million input tokens, output tokens free. The setup on this page stays identical.
Where does the full documentation live?
Vercel's knowledge-base guide 'How to classify, route, and score with Jev and AI SDK' covers everything on this page in more depth, and typesafe.ai has the original launch post.
The Chai Takeaway
Essays are expensive; multiple choice is cheap. Move one boring decision — your ticket router, your lead scorer — onto Jev this week, watch the confidence scores for a few days, and you'll understand the hype better than half the people arguing about it.
Read Vercel's full Jev guide