Jev pricing explained
Jev pricing explained is the compare frame: TypeSafe bills Jev like a decision function (input tokens, no generated prose), not like a chat completion (input plus expensive output). Snapshot date: 2026-09-21. List price and “output free” are vendor claims — docs.typesafe.ai/models wins if anything here drifts.
This page is for builders who already know LLM invoices and want the billing mental model before they sketch Jev pricing or a monthly planner. Product contrast (schema vs strings) lives on Jev vs an LLM.
Two meters, not one “AI tax”
A typical frontier LLM invoice has two axes TypeSafe’s launch table calls out (15 September 2026, their ranges):
- Input: they quote $0.20 to $10 / MTok for existing LLMs
- Output: they say output is “~5× more expensive than input tokens”
- Reasoning / chain-of-thought modes burn still more output tokens you never show the user
Jev 1.13, same launch table and the models page:
- Input: $0.042 / MTok ($42 per billion)
- Output: free (“too cheap to meter”)
- Answers are typed Choice / Score / Noul values, not a sampled essay
Jev still understands language. It does not emit a completion you pay per token. That is why the output meter can be $0 on the card — and why you cannot use Jev as a cheap writer.
Why “400× cheaper” is not a second price list
TypeSafe’s homepage-style multiples (they cite 193.6× faster and 444.6× cheaper on workflow evals) compare Jev to large, expensive generators doing the same workflow, often with structured wrappers. The launch post says those multiples sit at the high end of real-world gains and that evals were generally run from West Coast laptops.
Those claims are vendor-reported. They are not a discount off $0.042. Your gap versus your current model depends on:
- whether that model bills output (and reasoning tokens)
- how many questions you batch on one
state - gateway markup
- whether you were comparing to a flash-lite classifier or to a frontier agent session
Independent write-ups have measured single-digit to tens-of-times gaps against small LLMs, and much larger gaps against agent sessions. Cite those authors; do not flatten them into TypeSafe’s 444×. Jev review keeps methodology and third-party notes in one place.
What you pay for on a Jev request
You pay for input tokens: filtered state plus all question text (instructions, criteria, score levels) in that request. Docs: state is ingested once; questions evaluate in parallel against it.
You do not pay, on the 2026-09-21 card, for:
- the Choice key, Score level, or Noul float coming back
- a hidden chain-of-thought transcript (Jev does not generate one)
You may still pay a gateway for platform fees even when the model card repeats $0.042. Cloudflare’s public card said “view pricing in the dashboard” rather than repeating the sticker.
When the LLM invoice can still win
Use an LLM meter when you need strings: drafts, tool plans, explanations, long reasoning. JSON mode is still a generator with a format hope — see vs JSON mode.
Jev is the wrong bill to optimize if the job is “write the email.” It can be the right bill for “route the email, then let a generator write.” Code owns the handoff. Is Jev free? covers early access vs $0 fantasies.
FAQ
Why is Jev output free if LLMs charge for output?
Vendor explanation: there is no completion to meter. Confirm the models page; policies can change.
Does a cheaper input price mean Jev is “subsidized”?
TypeSafe wrote that they cannot prove the price is not subsidized and that they expect prices to go down. Treat that as their own caveat.
Should I compare Jev to input-only or to input+output LLM cost?
For a fair invoice compare, include the LLM’s output (and reasoning) tokens for the same decision. For a fair product compare, also ask whether you still need a generator after the decision.
Is this page a TCO calculator?
No. It is a billing mental model. Plug measured tokens into cost planning.
What this page does not claim
- No invented LLM prices beyond TypeSafe’s published comparison ranges.
- No jev.pro benchmark that “proves” 40×–400×.
- Schema-safe answers are not factual correctness — wrong labels still cost tokens.