Skip to content
Jevstiller

Jev savings calculator

Jev costs $0.042 per million input tokens and answers in about 300 ms. Jevstiller answers a repeated question locally once it has learned it from Jev’s own answers, within an agreement bound you set, and forwards the rest. This page estimates what that changes for one question at your volume. The local share comes from the benchmarks: five public tasks replayed against Jev’s recorded answers at a 98% target.

Jev today, per month

—

— requests, — input tokens, at $0.042 per million

With Jevstiller, per month

—

— saved (—). — of requests answered locally; the rest and a 2% audit still go to Jev.

Calls that never reach Jev

—

a month. They don't count against rate limits: Jev sees — of the traffic.

Time waiting on answers, per day

— → —

at a median of 300 ms from Jev and 16 ms from the local model, request by request.

  • Agreement: on this task the benchmark measured — agreement with Jev against the 98% target. The bound promises at most 2% of all requests answered differently from Jev, with 95% confidence. Agreement is not accuracy: where Jev is wrong, the local model is wrong the same way.
  • Warm-up: about 2,500 of this question's requests go to Jev before the first local answer (training, calibration and a shadow test), about — at your volume. They were going to Jev anyway.
  • These are replays, not your traffic. The local share comes from a public dataset replayed against Jev's recorded answers. Yours depends on how consistent Jev is on your question. Run the proxy in front of a copy of your traffic and read the status page before believing a number.

How the numbers are computed

Section titled “How the numbers are computed”
  • Jev today is requests × input tokens × $0.042 per million. Output tokens are free, so that is the whole bill for the question.
  • Answered locally is the task’s held-out coverage from the benchmark, times 98%: a random 2% of all traffic always goes to Jev as the audit that keeps the bound honest.
  • With Jevstiller is the same bill times the share that still reaches Jev. Jevstiller itself is free (Apache-2.0) and runs on one CPU container next to your services; that machine is the cost it adds.
  • Time waiting is the per-request median summed over a day: 300 ms for an answer from Jev, 16 ms for a local one, and 316 ms for a request the local model passes on, since it is encoded first.
  • Warm-up is 1,000 Jev answers to train the first model, 500 held out to calibrate the threshold, and 1,000 requests in shadow before it is promoted (configuration reference).

Every constant is at the top of the calculator’s source, each with where it comes from. The benchmark tables and the one-command reproduction are in benchmarks.

Before you deploy

Section titled “Before you deploy”

Jevstiller trains on Jev’s output, and TypeSafe’s terms restrict that. Read TypeSafe’s terms and decide for your own use. Then run the proxy in front of a copy of your traffic: the status page shows the real local share and agreement for your question.