Overview
Ask for the decision your application needs
Jev 1.14 is TypeSafe’s “System One” decision model. Instead of writing text, it answers the questions you define with typed results and calibrated probabilities: a yes/no probability, one of your choices or a level on your scale. It was released October 14, 2026 and is billed on input tokens only.
Define the contract
Put the content to evaluate in state, then add one or more questions. noul returns a yes/no probability, choice returns one of your keys with a probability for each, and score places the state on 2 to 10 ordered levels and returns the expected level. Answers always match the schema, so your code can branch on them without parsing free text.
Use the confidence
Every answer carries a confidence value. Act automatically on confident decisions and send uncertain ones to a person or a larger model. TypeSafe describes Jev as calibrated, meaning higher confidence goes with higher accuracy. Check that on a labelled sample of your own traffic before you choose a threshold.
What changed in 1.14
TypeSafe expects the largest gains in languages other than English. Version 1.14 also exposes image input and box questions, which locate an object in an image; TypeSafe marks both as experimental and not yet production-ready. Pin typesafe/jev-1.14 to keep behaviour fixed, because the typesafe/jev alias moves to each new release.
Benchmarks reported by
TypeSafe AI
These are TypeSafe’s own claims, with the scope TypeSafe gives them. We have not found an independent leaderboard that covers Jev.
What the claims cover
Read the numbers with their method before relying on them.
- Reference
- TypeSafe’s workflow evaluations score Jev against “the average of GPT-6 Astra and Fable 5.1” as the reference answer. The same post puts frontier LLMs at 3 to 329 seconds end to end.
- Range
- TypeSafe says the 193.6× and 444.6× figures are “on the higher end of real world gains”, and that because its own team built the workflows, “some bias could exist”.
- Type errors
- Answers always match the requested schema, so TypeSafe reports 0% by design. It is not an accuracy figure.
- Version 1.14
- The release post shows an internal intelligence measure that TypeSafe calls “likely biased towards our models”. We do not reproduce it here.
Sources: Introducing System One Models & Jev (September 15, 2026) and Jev 1.14 release notes, checked 2026-10-30. Figures describe Jev at launch, not 1.14 specifically, and were not measured by AI/ML API.
Pricing
| Usage | Price | Unit |
|---|---|---|
| Input | $0.042 | 1M tokens |
| Output | Free | typed decisions |
Compare decision model prices
AI/ML API input rates per million tokens. All three return decisions without an output charge.
| Model | Input per 1M tokens | Output |
|---|---|---|
| $0.042 | Free | |
| $0.042 | Free | |
| $0.042 | Free |
Jev Router is billed differently: the answering model’s rate plus one Jev decision.
Estimate by request volume
Assumes 500 input tokens per request. Long states and many questions cost more.
| Requests | Cost on AI/ML API |
|---|---|
| 1,000 requests | $0.021 |
| 100,000 requests | $2.10 |
| 1,000,000 requests | $21.00 |
| 10,000,000 requests | $210.00 |
Quick start
View docs (opens in a new tab)Build with your agent
Give your agent the model details, code and documentation.
1Get your API key
Create API key (opens in a new tab)export AIMLAPI_KEY="8f2b4c7d9a1e6b3f…"2Make your first request
curl 'https://api.aimlapi.com/v1/decisions' \
-H "Authorization: Bearer $AIMLAPI_KEY" \
-H 'Content-Type: application/json' \
-d '{
"model": "typesafe/jev-1.14",
"state": "I was charged twice for my subscription. Can you help me get a refund?",
"questions": {
"route": {
"type": "choice",
"instructions": "Choose the team that should handle this support request.",
"criteria": {
"billing": "Payments, charges, invoices and refunds.",
"technical": "Errors, account access and product functionality.",
"general": "General questions and product information."
}
}
}
}'import os
import requests
headers = {"Authorization": "Bearer " + os.environ["AIMLAPI_KEY"]}
payload = {
"model": "typesafe/jev-1.14",
"state": "I was charged twice for my subscription. Can you help me get a refund?",
"questions": {
"route": {
"type": "choice",
"instructions": "Choose the team that should handle this support request.",
"criteria": {
"billing": "Payments, charges, invoices and refunds.",
"technical": "Errors, account access and product functionality.",
"general": "General questions and product information."
}
}
}
}
response = requests.post(
"https://api.aimlapi.com/v1/decisions",
headers=headers,
json=payload,
timeout=120,
)
response.raise_for_status()
data = response.json()
answer = data["answers"]["route"]
print(answer["choice"], answer["confidence"])async function main() {
const headers = {
Authorization: "Bearer " + process.env.AIMLAPI_KEY,
"Content-Type": "application/json",
};
const response = await fetch(
"https://api.aimlapi.com/v1/decisions",
{
method: "POST",
headers,
body: JSON.stringify({
"model": "typesafe/jev-1.14",
"state": "I was charged twice for my subscription. Can you help me get a refund?",
"questions": {
"route": {
"type": "choice",
"instructions": "Choose the team that should handle this support request.",
"criteria": {
"billing": "Payments, charges, invoices and refunds.",
"technical": "Errors, account access and product functionality.",
"general": "General questions and product information."
}
}
}
}),
},
);
if (!response.ok) throw new Error(await response.text());
const data = await response.json();
const answer = data.answers.route;
console.log(answer.choice, answer.confidence);
}
main().catch(error => {
console.error(error);
process.exitCode = 1;
});Related articles
Frequently asked questions
How is Jev different from a chat model?
Jev returns typed decisions with probabilities, not natural-language text. Use it where your software needs a value to act on, such as a route, a label or a score.
Which question types are supported?
noul for a yes/no probability, choice for one of up to 255 keys you define, and score for an ordered scale of 2 to 10 levels. Version 1.14 adds box, an experimental question that locates something in an image.
How is it billed?
AI/ML API charges $0.042 per million input tokens; decisions have no output charge. At 500 input tokens per request, 1,000 decisions cost about $0.02.
Are image input and box questions ready for production?
Not yet. TypeSafe labels both experimental. Test them on your own images and keep a fallback in place.
Is Jev independently benchmarked?
We have not found an independent leaderboard for decision models that includes Jev. TypeSafe reports 70–500 ms responses and large speed and cost gains on its own workflow evaluations, and notes that some bias could exist. Measure accuracy on a labelled sample of your own data.
Why pin version 1.14?
typesafe/jev-1.14 stays on this release. The typesafe/jev alias follows the newest version, which can change the probabilities your application depends on.
How does Jev relate to Jev Router?
Jev Router uses Jev 1.14 to pick which model should answer a request, then forwards it. You pay the answering model’s rate plus one Jev decision.


