Jev: An AI Model That Makes Fast Choices

TL;DR

• Jev is an AI model from TypeSafe AI. It reads complex text and selects one answer from a list that you supply. It does not write text.
• Use Jev for many repeated decisions with fixed answers, for example lead scoring, policy checks, contract screening, and review tagging.
• Jev is 60 to 300 times cheaper than large LLMs on list price. 100,000 decisions cost $4.20 on Jev and $500 on Claude Opus 5.5.
• Jev does not replace LLMs. Use an LLM for writing and deep reasoning.
• Test Jev on your own data before you change models.

What Jev is

• Jev is an AI model from TypeSafe AI. TypeSafe released Jev on September 15, 2026.
• Jev does not write text. It reads text and selects one answer from a list that you supply.
• Jev also gives a confidence score for each answer.
• The technical name for this type of model is "classifier."
• TypeSafe calls Jev a "System One" model. System One is fast judgment. Large Language Models (LLMs), for example ChatGPT and Claude, do slow reasoning and writing.

How Jev works

1. Input: You send Jev complex text, for example an email, a support ticket, or a call transcript.
2. Question: You write a question and a list of answer options, for example "Is this customer at risk of leaving? Yes, No, or Unclear."
3. Decision: Jev reads the text and selects one answer. It also gives a confidence score.
4. Action: Your software uses the answer to do the next step, for example send the case to the retention team.

Why Jev is important

Before Jev, software had two tools:

1. Rules: "If an invoice is 30 days late, flag it." Rules cost little, but they cannot interpret meaning.
2. LLMs: LLMs can read, reason, and write. They are slower and cost more.

Many business questions need judgment but have a simple answer. For example: Does this customer want to cancel? Is this email a real sales lead? Does this contract need a lawyer? Companies often use an expensive LLM for these questions, or they do not ask them. Jev does this work faster and at a lower cost.

Speed, Cost, and Business Applications

Jev delivers large gains in speed and cost. TypeSafe reports that Jev runs up to 100 times faster than a large LLM and costs more than 100 times less. TypeSafe charges 4.2 cents per million input tokens and no charge for output tokens. 10,000 calls to Jev cost approximately 42 cents.

Businesses can apply Jev to tasks like these:

1. Lead Qualification: Jev reads web form submissions and inbound emails, decides if each lead is a good fit, and sends it to the correct sales rep.
2. Expense and Invoice Review: Jev reads expense reports and vendor invoices, and flags items that break company policy or look like duplicates.
3. Contract Screening: Jev reads incoming contracts and flags clauses that need legal review, for example automatic renewal or unlimited liability.
4. Customer Feedback Analysis: Jev reads product reviews and survey comments, and tags each one by topic (price, quality, service) and sentiment. 10,000 reviews of 500 tokens each cost approximately 21 cents.
5. Model Routing: Jev reads each request to an AI system and decides if a small, low cost LLM can do the task or if the task needs a large LLM. This keeps the expensive model for the hard work only.

Jev compared with LLMs

Function

Item	LLM	Jev
Output	Text	One answer from your list, plus a confidence score
Can write, explain, plan	Yes	No
Speed	Slower. Speed changes with answer length.	70 to 500 milliseconds, most near 100
Setup for a new task	Prompt	Question and answer options. No training.
Best use	Reasoning, writing, complex exceptions	Many repeated decisions with fixed answers
Main risk	Incorrect or inconsistent text	Incorrect answer with high confidence

List price per million tokens

Model	Input	Output
Jev	$0.042	No charge
GPT-6 Sol	$2.00	$10.00
Claude Opus 5.5	$4.00	$20.00
Claude Fable 5.1	$10.00	$50.00

Cost per 100,000 decisions (1,000 input tokens each; LLMs write 50 output tokens each)

Model	Cost	Jev cost difference
Jev	$4.20	Baseline
GPT-6 Sol	$250	Jev is approximately 60 times cheaper
Claude Opus 5.5	$500	Jev is approximately 120 times cheaper
Claude Fable 5.1	$1,250	Jev is approximately 300 times cheaper

Real tests

Test	Jev	Claude Opus 5	Difference
Support tickets: cost per 1,000	$0.0248	$2.88	Jev is approximately 116 times cheaper
Support tickets: accuracy (60 tickets)	59 of 60	60 of 60	Almost equal
Social post analysis, same 20 seconds	100,000 posts for $0.67	214 posts for $0.98	Jev is approximately 680 times cheaper per post

What the tables show:

• Against large LLMs, Jev is 60 to 300 times cheaper on list price.
• In real tests, the difference was 116 to 680 times.
• Large LLMs often use extra reasoning tokens, billed as output. This makes the real difference larger than the list price shows.
• In the ticket test, accuracy was almost equal. Test accuracy on your own data before you change models.

Limits

• Jev cannot write, explain, or plan. Use an LLM for these tasks.
• Jev can give an incorrect answer with high confidence.
• Bad input gives bad output.
• Jev is new. Independent test results are still limited.

Summary

Jev does not replace LLMs. It adds a third tool between rules and LLMs. Use Jev when you have many decisions with a fixed set of answers. Use an LLM when the task needs writing or deep reasoning.