Jev Pricing: What a Million Decisions Actually Costs

Jev is a System One decision model from TypeSafe. You send application state and typed questions, and it returns a typed answer with probabilities. On OpenRouter, typesafe/jev-1.13 costs $0.042 per million input tokens. Output is free, so the bill is set entirely by how much state and question text you send.

That pricing shape is unusual. With a chat model, a long answer can cost five times more per token than the prompt that produced it. With Jev, the answer is a label, a probability or a position on a scale, and it is free. The only thing you can overspend on is context.

Decision cost is just state tokens multiplied by the input rate. A compact ticket summary of about 500 tokens costs $0.000021 per decision, or $21 per million. A 2,000-token state such as a short email thread costs $84 per million decisions. A full 30,000-token state costs about $1.26 per thousand decisions, which is still cheap but sixty times the compact case.

Several questions about the same state can share one request. The OpenRouter tutorial routes a single ticket with three questions — is it a bug, which team owns it, how urgent is it — in one call. You pay for the shared state once rather than three times, so grouping questions is the most effective cost lever after trimming state.

Jev is billed to your OpenRouter account rather than a TypeSafe invoice, so OpenRouter's credit fee applies. The Standard plan charges 5.5% on credit purchases, which turns $21 of tokens into about $22.16. The self-serve Business plan launched in September charges 8%, about $22.68 for the same workload. In exchange it raises the workspace cap from 5 to 1,000 and adds US and EU in-region routing.

For most teams the plan decision will be driven by data residency and workspace count, not Jev. At these token volumes the gap between 5.5% and 8% is cents per million decisions. It matters more for the chat models Jev sits in front of.

The fair comparison is the prompt-and-parse pattern Jev replaces. Asking GPT-6 Luna to classify a 300-token item and return a 20-token label costs $30 of input and $10 of output per million items at $0.10 and $0.50 per million tokens: $40 in total. Jev answers the same 300-token question for $12.60 per million, with no parsing step and with per-option probabilities.

The cheap LLM still wins when the task needs generated text, a reason the user will read, or reasoning across a long document beyond Jev's 32,000-token window. Jev does not return explanations. OpenRouter's own guidance is to let Jev decide, and to call a chat model only when you need prose or when Jev's confidence is low.

The largest savings come from using Jev to avoid an expensive call, not from Jev's own low price. OpenRouter's cookbooks include a Jev-verified cascade: a cheap model drafts, Jev checks the draft against retrieved context, and only failed checks escalate to a stronger model. If 70% of drafts pass, 70% of requests never touch the premium model.

Tool-call gating works the same way. Jev can judge whether an agent's pending tool call matches the user's request, so safe calls run, unsupported calls are refused, and only ambiguous ones pause for a human. Each gate costs fractions of a cent and can prevent a costly agent loop or a destructive action.

Every Jev response carries a usage.cost field in US dollars. Log it next to the decision and the action your code took, so you can calculate cost per accepted decision rather than per call. Put Jev traffic in its own OpenRouter workspace and set a workspace budget. Workspace budgets arrived on every plan in September.

Jev reads instructions literally. Keep date arithmetic, threshold maths and joins in your own code, and send only the fields a question needs. That improves accuracy and lowers spend, because the state you strip out is the only thing you pay for.

Jev is not the only cheap way to answer a narrow question. In September OpenRouter also listed GPT-6 Luna at $0.10 and $0.50 per million tokens, Inception's Mercury 2.5 diffusion model at $0.04 and $0.15 during an 80% promotion, and DeepSeek V4.1 Flash at $0.30 and $1.20 with half-price off-peak hours. Each of these is a text model, so you still pay for the label it writes and for the code that parses it.

On a 300-token classification with a 20-token answer, Mercury 2.5 costs about $15 per million items at its promotional rate and GPT-6 Luna about $40. Jev costs $12.60 with no output charge and no parsing step. Jev also returns per-option probabilities you can threshold. When the promotion ends, or when answers need several fields, the gap widens in Jev's favour.

The comparison flips for open-ended work. If the label needs a sentence of justification for a human reviewer, a text model does both jobs in one call. Count the calls you would add before assuming Jev is cheaper end to end.