Jev pricing, and what a real build actually costs
Published Jev pricing is simple — input tokens are charged, output tokens are free — but the published input rate disagrees across sources by roughly 10x. Here are both figures, the per-decision math, and what builders in this directory report actually spending.
Updated September 2026 · 6 min read
The published numbers — both of them
Every Jev pricing page quotes the same headline: output tokens are free, you pay for input. They disagree on the input rate, and most do not mention that a second figure exists.
| Source | Input price | Output price |
|---|---|---|
| Launch post & docsquoted by most third-party write-ups | $0.042 per million input tokens | Free |
| Vendor marketing sitejevtypesafeai.com, read 2026-09-24 | $0.25–$0.42 per million input tokens | Free |
The two figures differ by roughly 10x, and we could not reconcile them from public material. Output tokens are free under either reading, so the practical difference lands entirely on input-heavy workloads.
What one decision costs
Per-million-token pricing is not the useful number. Decisions are the unit you actually buy, and TypeSafe publishes cost per case from their own workflow eval:
| Model | Cost / case | Time / case | Accuracy |
|---|---|---|---|
| Claude Opus 5 | $0.1761 | 37.8 s | 73.1% |
| Jev | $0.0004 | 0.07–0.5 s | 67.8% |
| Terra | $0.0304 | 10.1 s | 67.9% |
Reference answers in these evals came from two large models, not humans.
The ratio is the point: about $0.0004 per judgment against $0.1761 for the frontier model — call it three orders of magnitude. That is what makes asking a question at every loop iteration affordable instead of rationing it.
- Batch your questions. One call can carry many questions, and since output is free, the marginal cost of the second question in the same call is far below the first. Grouping is the single biggest lever you control.
- Input is the whole bill. Trim the state you send. A long transcript costs money; a distilled one does not.
- Free output changes retry economics. Asking twice is not 2x. It is close to 1x, because there is no second generation to pay for.
What builders in this directory report
Directory descriptions are written by the people who built the thing, and many of them state a cost. Extracting those gives a rough sense of the range real projects land in — not a benchmark, but a reality check on the per-case figures above.
Why output is free at all
Free output sounds like a promotion. It is actually a consequence of the architecture: Jev does not generate a token stream. It scores candidate answers, so there is no autoregressive decoding loop to bill for — the expensive part of serving an LLM simply does not exist here.
The practical upshot is a pricing shape you may not have seen before: cost scales with what you tell it, not with what it tells you. If your prompts are long and your answers are short — which describes most routing and classification workloads — the model is unusually cheap.
What actually drives your bill
- State length. The dominant term. Every extra paragraph of context is billed on every call.
- Call volume. Cheap calls invite more calls. Watch the count, not just the unit price — see the Jevons effect.
- Nothing about your answer length. A 255-way choice costs the same as a yes/no on the same state.
input tokens × calls. If you cannot estimate those two numbers, you cannot estimate your bill — and no per-million price will help.Common questions
How much does Jev cost per call?
Why is the input price different on different sites?
Are output tokens really free?
Is it cheaper to batch several questions into one call?
BuiltOnJev is an independent community project, not affiliated with TypeSafe AI. Specs and eval numbers come from TypeSafe's published materials; directory figures come from this site's own submissions and are updated as builds arrive.