Jev.

Most AI models answer questions with paragraphs. Jev does the opposite. Built by TypeSafe, it skips the prose and hands back a decision: a probability, a chosen option, or a score your code can act on straight away. If you have ever bolted a chatbot onto a workflow just to ask “does this support ticket need a human?”, then parsed a rambling paragraph to find the answer, Jev was built for that gap. It launched on 15 September 2026 as the first of TypeSafe’s “System One” models, in a browser playground and via API. This review covers its question types, what the credits and token packs cost, and where a decision-only model fits.

What Is Jev?

Jev is a decision model, not a chatbot. You send it a piece of text — TypeSafe calls this the “state” — together with typed questions about that text. Jev reads the state once, answers every question in parallel, and returns structured results with calibrated confidence, usually in about half a second. A yes/no question comes back as a probability; a choice question returns the winning option plus every option’s probability; a score question rates the text against levels you define.

The access site, jev-ai.pro, is an independent platform bringing TypeSafe’s Jev models into one workspace — playground, batch processing and API — and its current playground model is Jev 1.14. Because answers are numbers and labels rather than sentences, Jev fits the judgment calls inside larger systems: evaluating agent output, triaging tickets, matching entities, qualifying leads, scoring sales calls, routing MCP tool calls, or flagging prompt-injection attempts.

Key Features

• Three typed question types: Yes/No returns the probability the answer is yes; Choice picks one of your options and shows every option’s probability; Score rates the text on levels you set.
• Calibrated probabilities: Answers carry confidence values, so your code can branch on how sure the model is — and route borderline cases to a person.
• Many questions, one call: All questions in a request are answered in parallel against the same state, keeping multi-rule checks fast.
• API with a familiar shape: Send the same JSON (state, model, questions) you would send to TypeSafe, with batch processing for large volumes.
• Output tokens are free: Paid usage charges input tokens only, keeping costs predictable on question-heavy workloads.

How to Use Jev

1. Create an account on jev-ai.pro — eligible new accounts receive 20 welcome credits.
2. Pick a case in the playground, or paste your own text (plain text or JSON) as the state.
3. Write your questions and pick a type for each: Yes/No for binary checks, Choice for classification, Score for graded judgments.
4. Run Jev and read the probabilities, adjusting questions and thresholds until the results match your own judgment.
5. Ship it with the API — generate an API key and send the same request structure from your application.

Jev Pricing

Jev runs on two balances. The Credit account holds your 20 welcome credits (plus daily check-in gifts) and is spent first by the web tools, at 1 credit per run on the standard model. The Token account holds purchased input tokens, which the API spends first. A request charges only one account, output tokens are always free, and failed requests are not charged. Subscriptions also exist (annual billing is 50% off), but one-time packs are simpler — prices verified on the official pricing page:

Plan	Price	What you get
Start	Free	20 welcome credits for eligible new accounts, plus daily check-in gifts
15M Token Pack	$9	15M input tokens one time (about $0.600 per 1M); tokens never expire, no subscription needed
75M Token Pack	$39	75M input tokens one time (about $0.520 per 1M); unused tokens roll over
200M Token Pack	$76	200M input tokens one time (about $0.380 per 1M) for regular playground and API use
600M Token Pack	$170	600M input tokens one time (about $0.283 per 1M), the largest pack, with 20% extra tokens included

Note that bigger platform models cost more per run: Jev runs at a 1× rate, Clef at 6× and the multimodal Jev-Omni at 4×.

Pros and Cons

Pros

• Returns decisions your code can branch on immediately — no parsing paragraphs to find a “yes”
• Calibrated confidence makes review rules simple: auto-accept sure cases, escalate the rest
• Roughly half a second per call, all questions answered in parallel
• Input-only billing and free output keep question-heavy workloads predictable
• 20 welcome credits and non-expiring token packs make evaluation cheap

Cons

• The core model reads text only — no PDF support, and images or audio need the pricier Jev-Omni model
• English is the primary language; the makers say other languages work less reliably
• The two-balance credit/token system takes a little reading before it clicks
• A young platform, so the ecosystem and community are still small
• By design it cannot draft or explain — if you need prose, you need another model alongside it

Who Should Use Jev?

Jev suits developers and teams whose AI problem is judgment rather than generation. If you run an agent pipeline and need a cheap, fast checker to grade outputs, route requests between tools, screen text for injection attacks, or decide which tickets and leads deserve human attention first, typed answers beat prompting a general chatbot and hoping the format holds. It also works well as an evaluation layer, scoring model answers before they reach users. For writing or conversation, look elsewhere — Jev is the referee, not the player. Compare it with the other automation tools in our AI tools directory.

Frequently Asked Questions

Is Jev free to use?

Eligible new accounts get 20 welcome credits plus daily check-in gifts, so testing costs nothing. Beyond that, usage runs on paid token packs or a subscription.

What kinds of questions can Jev answer?

Yes/No questions return a probability; Choice questions pick one option and show every probability; Score questions rate text on levels you define. All are answered in parallel.

How is Jev different from ChatGPT or Claude?

Those models reply in prose. Jev returns structured decisions — probabilities, labels and scores — with calibrated confidence, in about half a second. It sits inside software workflows rather than holding conversations.

Can Jev read images or PDFs?

Native Jev accepts text only — a string, JSON object or array — and PDFs are not supported. For images, audio and short videos, the playground offers Jev-Omni at a higher rate.

How does the Jev API work?

Create an API key, then send a JSON request with your state, model and questions — the same TypeSafe-compatible shape used in the playground. Failed requests are not charged.

Final Verdict

Jev is one of the more interesting ideas in AI tooling this year: a small, fast model that does one job — judging text — and tells you how confident it is. For agent builders, support teams and anyone tired of parsing chatbot prose to extract a decision, its typed answers, parallel questions and input-only pricing add up to a practical tool. The limits are real (text-first, English-first, young platform), but the welcome credits cost nothing — the easiest way to judge Jev is to let it judge something of yours.

Visit Jev