Jev pricing: what a decision actually costs

Input

$0.042 per 1M input tokens

Output

Free (no output tokens are generated)

Latency

70–500ms end-to-end

Context

64k tokens

Rate limits

100,000 tokens per second and 80 requests per second (TypeSafe says these are adjusting and can change without notice); either one over the line returns 429

Free tier

No free-credit or trial program documented at launch

Measured, not estimated

These are real calls against the live API on 2026-09-18, with the cost as the API itself reported it — not arithmetic from the rate card. One caveat worth stating: that cost field comes from OpenRouter. The first-party API returns input_tokens and output_tokensbut no cost, so going direct means computing spend from tokens yourself:

Why is output free?

Because Jev does not generate any. A chat model bills you per token it writes; Jev returns a decision and a probability distribution from a single forward pass. There is no stream of tokens to meter, so the entire cost sits on the input side.

The most one call can cost

Because Jev pricing is input-only, a single request has a hard ceiling. The context window, 64k tokens, covers the state plus every question, and nothing billable is generated on top of it, so even a request filled to the limit costs under $0.003 at the rate card. Spend grows with volume and with state size, never with a model deciding to say more.

A worked example

Jev pricing is easiest to see on a real workload. Say you are routing support tickets: an average ticket plus your option list is around 500 tokens of state. At $0.042 per 1m input tokens:

500 tokens x 1,000,000 tickets  = 500,000,000 input tokens
500M tokens / 1M x $0.042       = $21.00

One million routing decisions ≈ $21

The same job done by asking a frontier chat model to emit JSON costs orders of magnitude more, runs in seconds rather than milliseconds, and needs parsing plus retry logic you have to maintain. That gap — not the raw number — is the pitch.

Jev pricing on each route

Jev pricing is the same on all three routes: TypeSafe, OpenRouter and the Vercel AI Gateway list $0.042 per million input tokens and nothing for output. What differs between routes is who meters you and whose terms apply.

• TypeSafe AI (first-party) — Official API. Get a key from the console dashboard; check your account for current availability.
• OpenRouter — Live as `typesafe/jev-1.13` at $0.042/M input, $0 output, 32K context, ~99.8% availability. Note it will not show up in OpenRouter's default model list — see the gotcha below.
• Vercel AI Gateway — Also carries it, as model id `typesafe-ai/jev`, at the same $0.042/M input, $0 output.
• Cloudflare AI Gateway — Announced by Cloudflare as live and callable from Workers. Third-party announcement rather than TypeSafe documentation — verify the model id against Cloudflare's own catalogue before you build on it.

Gateway listings matched the first-party rate at our last check, but a gateway can add a margin later, and its limits and terms apply instead of TypeSafe's. Check both before committing production traffic.

No TypeSafe key yet? This site sells prepaid credits for its own endpoint, which forwards to Jev: one credit per started 1,000 input tokens of a request. That is a separate price list from the Jev pricing above, and it is on Plans.

Is self-hosting cheaper?

Rarely, at today's Jev pricing. A 4B open model on your own GPU has a real hourly floor whether or not you are serving traffic, plus the engineering to build and calibrate it. At $21 per million decisions, the hosted API is hard to beat on cost alone — self-hosting makes sense when you need data residency, offline operation or no vendor dependency. See the open-source options if that is you.

For an hourly number to hold against a GPU, TypeSafe gives one for its own Doom bot: about $7 an hour at roughly ten queries a second. That is the vendor's figure, not ours.