Jev Cost Per Request: The Napkin Math With Real Receipts

Nobody publishes a trustworthy single number for "cost per request," including TypeSafe, whose official pricing is the only binding source and can change. What builders publish instead is better: whole workloads with totals attached, from which per-verdict cost falls out by division. This page does the division.

The receipts, as reported by their authors: 500 email classifications for ~3.5 cents works out to roughly seven-thousandths of a cent per verdict. ~26,000 judge verdicts for $0.13: about five-thousandths of a cent each. The SuperX scorer's 61 questions for $0.0004: under a thousandth of a cent per question. Sustained high-frequency use, Doom at ~10 queries/second for ~$7/hour, implies a couple hundredths of a cent per decision at game tempo. Independent builders, unrelated workloads, and the derived per-verdict figure keeps landing in the same band: thousandths to hundredths of a cent, varying with question and context size.

The napkin method for your pipeline, from the pricing page: items per day × questions per item × runs per item, multiplied by the closest reference band above. A 2,000-ticket day at five questions each prices in dimes; a full eval suite per commit in pennies; only always-on real-time loops earn dollars. Which is why the standing advice is unromantic: meter from day one anyway, because at these prices the risk was never the rate, it's the retry loop nobody noticed for a month.

Frequently asked questions

So what's the per-request price?

Derived from reported workloads: thousandths to hundredths of a cent per verdict, question-size dependent. Binding rates are TypeSafe's docs, not derivations, ours included.

Do longer questions cost more?

Reported figures suggest context size moves cost within the band; decomposed short questions per the craft guide stay at the cheap end and judge better anyway.

At what scale does Jev stop being cheap?

Continuous high-frequency loops (real-time, per-tick) reach whole dollars per hour; batch and triage workloads essentially never leave pocket change at the volumes most products see.