Jev Pricing Explained

If you are used to LLM billing, Jev's pricing looks wrong at first glance. There is no output charge at all. Here is how it actually works, what a real bill looks like, and where the costs that never appear on the model invoice go.

Numbers below are for Jev 1.13 from the official models page. Check that page before you budget — TypeSafe has been adjusting things during the launch period.

You pay for input only

Jev bills per input token at $0.042 per million ($42 per billion). Output tokens are free. Every answer, including the choice, the full probability distribution, and the confidence value, costs nothing.

A few things fall out of that. Your bill is a function of how much text you send, and since states and questions are yours, you can work out the cost of a decision before you make it. There is no reason to fear a runaway answer, because Jev cannot produce one — it returns typed values and stops. And the usual pressure to compress prompts mostly disappears. At $0.042 per million tokens you trim context when it improves accuracy, not to save money.

What a real bill looks like

monthly cost =
  states per day × avg input tokens × 30 days × $0.042 / 1,000,000

Say you run a support router over every inbound ticket. Each state is about 400 tokens (the ticket plus a policy excerpt) and carries one Choice question.

10,000 tickets/day × 400 tokens × 30 days  =  120M input tokens/month
120M × $0.042 / 1M                          ≈  $5.04 / month

That is the entire AI cost of routing a support queue. Run the same arithmetic for each decision you plan to move: agent routing, tool selection, guardrail checks. Most products have several, and an agent loop that gates every step can make three to five decisions per user turn. Watch the token counts as well as the decision count — a 4,000-token agent transcript costs ten times a 400-token ticket.

Where it adds up

The other side of the ledger matters too.

Fat states are the real cost driver. Pasting whole documents or transcripts into every decision multiplies tokens linearly. Jev's context window is 64k per request, which is generous, but you pay for what you fill.

High-volume loops hit rate limits before they hit budget. Jev 1.13 allows 250k tokens per second and 1,200 requests per minute, and TypeSafe says those limits are moving during the launch period. At 10,000 tickets a day you are nowhere near them. At 10,000 decisions a minute you need sharding or an enterprise agreement.

Escalations still bill somewhere. Low-confidence decisions that get handed to an LLM are LLM costs, so budget both legs of the workflow.

The costs that never appear on a model invoice

The sticker price is only part of the comparison. Moving a classifier off an LLM also removes work that was never billed per token.

Parsing and repair. LLM structured output fails sometimes, and someone pays for the retries, the fallback path, and the incident. Jev returns typed answers by contract.

Latency on the hot path. A slow classification blocks whatever sits behind it. That shows up in infrastructure sizing, timeouts, and user churn rather than in the AI bill.

Prompt maintenance. Keeping a classifier prompt stable across model versions is real engineering time. instructions and criteria configs drift far less.

When an LLM is still the cheaper answer

At small volume the price barely matters; a few hundred decisions a month costs cents either way. The honest comparison, which we go through in Jev vs LLM, is that Jev's advantage is the combination of latency, calibrated confidence, and typed output at a price that stops being an argument. If a decision runs once a week and needs heavy reasoning over a long document, keep the LLM.

Budget checklist

List every decision point you plan to move: routing, tool selection, guardrails
Multiply decisions per request by requests per day — agent loops multiply fast
Estimate input tokens per decision (state plus questions). This is your only billed usage
Apply the current rate from the official models page
Check your requests per minute against the current limits
Add the escalation path, because low-confidence decisions still hit an LLM

Test before you estimate

A spreadsheet is only as good as its assumptions. The Jev Agent playground lets you run your own inputs through routing and tool-selection scenarios, and the official TypeSafe playground shows the raw primitive results. Get your numbers from observation instead of guessing.