BuiltOnJev

Jev statistics: latency, cost, and this directory's own numbers

Benchmark numbers for Jev are scattered across launch posts, docs and dozens of re-writes, usually without the caveat that makes them interpretable. This page keeps them in one place, adds the caveat, and pairs them with numbers nobody else has: our own directory.

Updated September 2026 · 5 min read

This directory, right now

These come from our own database and move as listings and bids arrive. They are not a benchmark of Jev — they are a snapshot of what people have built with it and how the community responded.

MetricValue
Builds listed45
Community upvotes cast2
Outbound clicks tracked45
Spotlight bids placed0
Cumulative spotlight spend—

Spotlight bids start at $5 in steps of $5. Directory rank is upvotes only and never affected by spend.

One honest reading of that table: clicks and upvotes are the numbers that say whether a listing found an audience, and the spotlight spend is the only figure money touches. We publish both separately on purpose.

Published performance

TypeSafe's own workflow eval, reproduced here with the caveat that most re-posts drop:

ModelAccuracyCost / caseTime / case
Claude Opus 573.1%$0.176137.8 s
Terra67.9%$0.030410.1 s
Jev67.8%$0.00040.07–0.5 s
Claude Haiku 4.553.6%——

Reference answers in these evals came from two large models, not humans.

The headline claims that get repeated are 20–200x faster than an LLM and 40–400x cheaper than an LLM; output tokens are free. Both come from this same eval and inherit the same caveat about non-human reference answers.

Specs, and where the numbers disagree

SpecValueNote
Modelsjev-latest, jev-preview—
Latency70–500 ms per callPer call, typically
Max choice options255Per question
Context window64k tokenssome third parties say 32k tokensIf your workload depends on the exact ceiling, check the vendor's model page before designing around it.
Rate limits250k tokens/s, 1,200 requests/minTypeSafe's published materials
Input price$0.042 per million input tokensvendor site states $0.25–$0.42 per million input tokensOutput tokens are free under either reading.
We flag rather than resolve
Where public sources disagree we show both figures instead of picking the flattering one. If a number here matters to your design, check it against TypeSafe's own documentation before committing.

How to read these numbers

  • Accuracy is the weak axis. Jev is not more accurate than a frontier model — it is faster, cheaper and type-safe. Optimise for the axis you actually need.
  • The eval is not yours. A vendor workflow eval with model-generated reference answers tells you roughly where a model sits, not how it will do on your distribution.
  • Latency is a distribution, not a constant. 70–500 ms per call is typical; the tail matters more when you are putting it inside a loop.
  • Our directory numbers measure interest, not quality. A listing with many clicks found an audience. That is a different claim from being good.

Common questions

How accurate is Jev?
TypeSafe's published eval puts it at 67.8% on their workflow benchmark, against 73.1% for a frontier LLM. Reference answers in that eval came from two large models rather than humans.
How fast is Jev?
70–500 ms per call per call typically, against tens of seconds for a frontier model on the same judgment. It scores candidate answers rather than generating a token stream.
How many builds are listed here?
The live figure is in the table at the top of this page and updates as submissions arrive. Every listing passed a Jev review for Jev-ness before it counted.
Why do different sites give different Jev prices?
Two figures are in circulation: $0.042 per million input tokens in the launch post and docs, and $0.25-$0.42 per million on the vendor's marketing site. We could not reconcile them, so we report both.

BuiltOnJev is an independent community project, not affiliated with TypeSafe AI. Specs and eval numbers come from TypeSafe's published materials; directory figures come from this site's own submissions and are updated as builds arrive.