Jev, the decision model from TypeSafe AI, has one of the simplest pricing pages in the industry because it has one line: $0.042 per million input tokens, output free. No tiers, no free allowance, no per-seat anything. That simplicity is worth unpacking, because the number only becomes meaningful when you work out what a decision actually costs.

What you pay for

A Jev call has two parts: the state (the text you want judged) and the questions (what you want to know about it). Both are input. What comes back — typed answers with probabilities and confidence — is output, and output is free. This is the inversion that makes the price work: with a chat model you pay for the words it generates, so every token of a verbose answer is billable. With Jev the answer is structured data, and TypeSafe charges only for the reading.

Worked examples

A realistic support-ticket triage call is small. Say the ticket is 200 tokens, your instructions and choice criteria add 300 more. At $0.042 per million tokens:

  • One call costs 500 × 0.000042 / 1000 = $0.000021 — about two thousandths of a cent.
  • Ten thousand tickets a day at that size cost $0.21 per day.
  • A million calls a month cost $21 per month.

Compare that with routing the same ticket through a chat completion. You need the same input tokens, but the model must also write an answer you then parse. Output tokens are typically billed several times higher than input on small chat models, and a single malformed answer that crashes your parser silently costs a retry. The vendor claims Jev is 40–400× cheaper for decision work than chat models; the direction of that claim matches the arithmetic above, and every build in [this directory](/) that reports both numbers got there the same way.

The parallel-questions multiplier

The part of Jev pricing that surprises people is not the per-token rate but the parallelism. All questions in one call are answered against the same state and billed once. Sixty questions about one document cost about the same as one question, because the state — the expensive part — is read a single time.

That changes how you design. Instead of six API calls with six prompts, you send one state and a question block: which team owns this, is it urgent, is it a refund, should we escalate, what language is it, is it polite. Same input cost, six typed answers. The choice, score, and noul guide covers what each question type returns.

Where the bill actually grows

In practice, three things drive cost, and none of them is the answer:

1. State size. Pasting an entire document when a paragraph would do. Trimming the state is the only lever that matters at scale.

2. Call count. Batching related questions into one call instead of several.

3. Redundant re-judging. Caching a verdict for an unchanged input instead of re-reading it.

Latency behaves the same way: TypeSafe reports 70–500 ms per call regardless of how many questions you ask, which is what lets Jev sit inside a request path — routing, moderation, ranking — rather than behind a job queue.

What pricing does not tell you

Two honest caveats. First, there is no free tier, so the first dollar is yours; at these rates it buys millions of decisions, but you cannot experiment at zero cost the way you can with some chat APIs. Second, the numbers above are arithmetic on the published rate, not a benchmark. When you want real-world costs, the entries in this directory that carry a reported cost show what builders measured on their own workloads — and those receipts, not vendor math, are the ones worth quoting.

For what the model actually does with your tokens, start with What is Jev. For the same job priced both ways, see Jev vs an LLM.

Published 25 September 2026. Pricing checked against TypeSafe AI’s published rate on this date; worked examples are arithmetic, not measurements.