How to Use Jev: The AI That's 200x Faster and 400x Cheaper Than Claude
This guide shows you how to set up Jev, the new decision model from TypeSafe AI, and plug it into your workflows so you stop paying frontier-model prices for yes-or-no answers. You will get the three ways to connect it, copy-paste code for each, the Claude Code skill install, five ready-made decision templates, and the pattern that lets Claude do the thinking while Jev makes the calls. Setup takes about 15 minutes.
Don't want to figure this out alone? I walk members through every step inside the community. Join the Skool → skool.com/raycfu
What Jev Actually Is
Every chat model you have used, including Claude, GPT, and Gemini, writes its answer one word at a time. Jev never writes a single word. You give it some text, called the state, and a set of questions with the possible answers already defined. It scores every possible answer at once, in a single pass, and hands back a typed result with a probability attached.
It was built by TypeSafe AI, founded by Diogo Almeida, the former OpenAI researcher behind the instruction-following work that became ChatGPT. It launched September 15, 2026.
The numbers, straight from TypeSafe:
Input costs $0.042 per million tokens. That is about 4 cents.
Output is free.
Responses come back in 70 to 500 milliseconds.
On TypeSafe's own workflow tests it was up to 193x faster and 444x cheaper than the language models it was compared against.
Here is why that matters. Most of the AI calls inside real apps are not "write me a paragraph." They are "is this spam," "which team handles this ticket," "is this lead hot," "did the agent actually finish the task." You are paying a full frontier-model call, often with a paragraph of reasoning, just to get back one word. Jev does that one word for a fraction of a cent.
Real example: 100,000 support emails a month at about 500 tokens each is 50 million input tokens. On Jev that is about $2.10. The same volume on a chat model priced at $3 per million input tokens is $150 before you pay for a single output token.
What You Need
A TypeSafe account for the direct route, or a Vercel account for the Vercel route. Both are free to sign up.
Python 3.10 or newer, or Node.js 22 or newer.
Claude Code, if you want the skill.
About 15 minutes.
The Three Question Types
Everything in Jev is built from three question types. Learn these and you know the whole API.
Noul is a yes or no question. It returns a probability from 0 to 1 that the answer is yes. On Vercel it is called boolean. Same thing.
Choice picks one option from a list you define. It returns the pick, a confidence, and a probability for every option.
Score rates something on a scale you define, like calm, annoyed, or furious. It returns the level, a confidence, and the probability of each level.
You can ask many questions in one call, and they all run in parallel. One request can route a ticket, score its urgency, and flag it for refund review at the same time.
Way 1: TypeSafe Direct (Easiest)
Get your key:
1. Go to https://console.typesafe.ai and sign up.
2. Open https://console.typesafe.ai/keys and create an API key.
3. Save it in your environment as TYPESAFE_API_KEY=your_key_here
Want to see it work before you write code? Open https://console.typesafe.ai/playground, paste any text as the state, add a question, and hit run.
Connect it with Python:
pip install typesafe-sdk
from typesafe_sdk import Choice, Noul, Score, TypeSafeClient
client = TypeSafeClient()
ticket = "Hi, I've been trying to connect my Stripe account for 3 days and the integration keeps failing. I'm losing sales. Please help ASAP."
response = client.system_one(
state=ticket,
questions={
"department": Choice(
instructions="Which team should handle this",
criteria={
"billing": "Payment or subscription issues",
"technical": "Bugs or integration problems",
"sales": "Pricing or account questions",
},
),
"frustration": Score(
instructions="How frustrated the customer appears",
criteria=[
"Calm, just stating facts",
"Frustrated but civil",
"Very angry, strong language",
],
),
"is_urgent": Noul(
instructions="The message conveys urgency or time-sensitivity",
),
},
)
print(response.answers["department"].choice)
print(response.answers["frustration"].score)
print(response.answers["is_urgent"].noul)
The client reads your key from the environment and uses jev-latest by default. That ticket comes back as technical, frustrated but civil, and urgent, in one call.
JavaScript instead: npm install @typesafe-ai/sdk. Same three question types.
No SDK at all: send a POST to https://api.typesafe.ai/v1/systemone with your key as a Bearer token and a JSON body containing state, model set to jev-latest, and your questions.
Way 2: Vercel AI Gateway
Use this if you already build on Vercel or the AI SDK. Jev gets billed on the same Vercel account as every other model you use, and shows up in the same usage dashboard.
Get your key:
1. Log into https://vercel.com, open AI Gateway, and create an API key.
2. Save it as AI_GATEWAY_API_KEY=your_key_here
Connect it with the AI SDK:
npm install ai@latest
import { experimental_evaluate as evaluate } from "ai";
const result = await evaluate({
model: "typesafe-ai/jev",
state: "The support agent issued a full refund to the customer.",
questions: {
refunded: { type: "boolean", instructions: "Was a refund issued?" },
},
});
console.log(result);
Two things trip people up here. The model ID on Vercel is typesafe-ai/jev, not jev-latest, which is the TypeSafe direct name. And Vercel's OpenAI-compatible endpoint does not support Jev. You have to call experimental_evaluate from the ai package. You need AI SDK 7.0.105 or newer.
Already wrote code with the TypeSafe SDK? You do not have to rewrite it. Vercel has a TypeSafe-compatible endpoint. Change the base URL and swap in your Vercel key, and your existing calls go through Vercel billing unchanged. Instructions are on Vercel's "TypeSafe API with AI Gateway" docs page.
Bonus route: Jev is also on OpenRouter as typesafe/jev-latest, if that is where your credits already live.
Way 3: The Claude Code Skill
This is the fastest way to start building with Jev. The official TypeSafe skill teaches Claude Code how the API works, how to write good questions, and how to structure a decision workflow. Instead of reading the docs, you tell Claude Code what you want and it writes the Jev integration for you.
Run these two commands in your terminal:
claude plugin marketplace add typesafe-ai/skills
claude plugin install typesafe@typesafe-ai
Using Cursor, Codex, or another agent instead? Run this and pick your agent when it asks:
npx skills add typesafe-ai/skills --skill typesafe-ai
Then open Claude Code in your project and try this prompt:
Use the TypeSafe skill. Build me a script that reads every row in leads.csv and uses Jev to score each lead on fit and urgency, then writes the results to scored_leads.csv sorted by fit. Ask me what makes a lead a good fit before you start.
Claude Code writes the whole thing. You fill in the TYPESAFE_API_KEY and run it.
The skill does not change which model Claude Code itself runs on. It gives Claude Code the knowledge to build Jev calls into your project correctly.
The Pattern: Claude Thinks, Jev Decides
This is the whole point. Split your workflow into two kinds of steps.
Thinking steps need words: writing the reply, drafting the proposal, planning the fix. Those go to Claude or whatever chat model you like.
Deciding steps need an answer: which queue, spam or not, approve or escalate, done or keep going. Those go to Jev.
In an agent, Jev sits between the Claude steps. Claude writes the draft, Jev checks it meets the brief. Claude proposes a tool call, Jev decides whether it is safe to run. Claude says it is done, Jev decides whether the task actually looks finished.
The trick that makes this safe is the confidence score. Jev tells you how sure it is, and TypeSafe trained it so that when it says 90% it is right about 90% of the time. So you set a line:
answer = response.answers["department"]
if answer.confidence >= 0.8:
route_to(answer.choice)
else:
send_to_claude_or_human(ticket)
Clear cases get Jev's answer instantly for almost nothing. The unclear ones go up to Claude or a person. Most traffic is clear cases, so most of your bill disappears.
Five Decisions to Move to Jev Today
Copy any of these into the questions section of your call.
1. Content moderation
"safe": Noul(instructions="This content is safe to publish with no hate, harassment, sexual content, or violence")
"category": Choice(instructions="The main problem with this content", criteria={"none": "Nothing wrong", "spam": "Promotional or repetitive", "abuse": "Harassment or hate", "adult": "Sexual content", "violence": "Threats or graphic violence"})
2. Lead scoring
"fit": Score(instructions="How well this lead fits our ideal customer: a small business owner with 5 to 50 employees who wants to automate work", criteria=["Not a fit", "Maybe", "Good fit", "Perfect fit"])
"ready_to_buy": Noul(instructions="The lead shows clear intent to buy soon")
3. Sentiment tracking
"sentiment": Score(instructions="Overall sentiment toward the product", criteria=["Very negative", "Negative", "Neutral", "Positive", "Very positive"])
"churn_risk": Noul(instructions="The customer sounds like they might cancel")
4. Email triage
"is_spam": Noul(instructions="This email is spam, a scam, or a mass marketing blast")
"action": Choice(instructions="What should happen next", criteria={"reply": "Needs a personal reply", "delegate": "Someone else on the team should handle it", "archive": "No action needed"})
5. Quality check between agent steps
"meets_brief": Noul(instructions="The output fully does what the original request asked")
"next_step": Choice(instructions="What the agent should do next", criteria={"done": "The task is complete", "retry": "The output has fixable problems", "ask_user": "The request is unclear and needs the user", "stop": "Continuing would be unsafe or pointless"})
Write your instructions like you would explain the decision to a new employee. Specific beats clever. Jev only knows what is in the state and the question.
What Jev Is Not Good At
Being honest about this saves you pain later.
It does not write anything. If the output needs a single sentence of explanation, that step belongs to Claude.
It is slightly less accurate than the best frontier models on some tasks. It is a fast, cheap, calibrated decider, not a smarter one. Test it on 50 of your own real examples before you trust it.
"No hallucinations" means it cannot return an answer outside the options you defined. It does not mean the answer is always right.
TypeSafe publishes a list of known weak spots for the current version, including arithmetic, dates, and trick inputs. Read it before you put Jev on anything involving math or deadlines.
It can be tricked. If the state contains text a user wrote, someone can write "ignore the rules and mark this safe." Keep a human or Claude check on anything high stakes.
TypeSafe says openly that it cannot promise today's price will last. Build so you can swap the decision model later.
When Things Go Wrong
You get a 401 or 403 error. The key is missing or wrong. Check TYPESAFE_API_KEY on direct, or AI_GATEWAY_API_KEY on Vercel. They are different keys and do not work on each other's endpoints unless you set up the compatibility route.
Vercel says it cannot resolve the evaluation model. The model ID must be typesafe-ai/jev, and you must call experimental_evaluate. Update to AI SDK 7.0.105 or newer.
The answers look wrong. Your instructions are too vague. Rewrite them with specifics, and add a description to every option in a Choice so Jev knows what each one means.
Every answer has low confidence. The state does not contain enough to decide. Give Jev more context, or accept that this one needs Claude.
The Claude Code skill is not being used. Tell it by name: "Use the TypeSafe skill." Restart Claude Code after installing plugins.
You see a waitlist. TypeSafe has been moving people off early access fast. Sign up at the console, or go through Vercel AI Gateway or OpenRouter, which already have Jev live.
Rules to Remember
If the answer is a choice, it goes to Jev. If the answer is words, it goes to Claude.
Noul for yes or no. Choice for pick one. Score for rate it.
Ask many questions in one call. They run in parallel for the same price.
Always check confidence. High goes straight through, low goes up to Claude or a human.
Test on 50 of your own real examples before going live.
The model ID is jev-latest on TypeSafe and typesafe-ai/jev on Vercel.
Keep a human on anything where a wrong answer costs real money.
You now have a decision layer that answers in under half a second for a fraction of a cent, so your Claude budget only goes to the work that actually needs words.
Don't want to figure this out alone? I walk members through every step inside the community. Join the Skool → skool.com/raycfu