Models & Pricing: jev-1.13.0, Costs, Limits & Honest Benchmarks

Everything currently known about the Jev model lineup, pricing, and operating limits — with the accuracy caveats up front where they belong, not buried in a footnote.

#Model IDs

Rule of thumb: develop against jev-latest if you like living dangerously, pin jev-1.13.0 (or whatever the current pinned release is) for anything users depend on, and re-run your calibration set when you bump the pin.

#Pricing

Output being free is the pricing tell of the whole paradigm: Jev's outputs are tiny typed verdicts, so TypeSafe AI can afford to not meter them. The practical consequence is that your cost scales with state size and nothing else — another reason State Design (filter first, send records not documents) is a cost discipline as much as an accuracy one. A lean 300-token state costs about $0.0000126 per question batch; even a million such calls a day is ~$12.60/day before you add questions.

#Limits & performance

#The honest accuracy caveats — read these before trusting any number

This is the part most launch coverage skips, and the part we consider mandatory:

1. Independent re-tests have landed below official accuracy claims. How far below varies by task — sometimes a little, sometimes a lot.
2. Accuracy is jagged. Jev can be excellent on one framing of a task and mediocre on a rephrased version of the same task. A single benchmark number — ours, theirs, anyone's — does not transfer to your workload.
3. One early third-party benchmark was withdrawn. Treat the entire first wave of "Jev scored X%" coverage as suspect unless you can find the methodology and it survived scrutiny.
4. No SLA on latency, and rate limits may change during early access. Don't put Jev on a hard real-time critical path without a fallback.

None of this makes Jev useless — cheap, fast, approximately calibrated judgment at volume is valuable even at "good but not magical" accuracy. It makes measuring on your own data (see Confidence & Calibration) a required step, not a best practice.

#Cost math for the front-filter pattern

A concrete comparison, using list prices as of 2026-09-20: filtering 1M emails with a 400-token state costs ~$16.80 in Jev input. Running the same items through a frontier LLM at even $1/M input tokens costs ~$400 — before output tokens, which Jev doesn't charge and LLMs do. That ~25–100× gap (demos claim more; treat multipliers from demos skeptically) is the entire economic argument for Cascade Routing: Jev everywhere, LLMs only where generation pays for itself.

#Where to go next

• API Reference — endpoint, fields, errors
• Getting Started — make the first call
• Cascade Routing — spend Jev calls to save LLM calls

#Sources

• jevai.wiki — Model card and API reference (community documentation; model IDs, pricing, limits).
• learnjev.com — Cost & benchmarks (community documentation; independent re-test results and the withdrawn-benchmark story).
• jev101.com — 什么是 Jev（中文） (community documentation, Chinese).

Unofficial fan-made handbook. Not affiliated with TypeSafe AI or jev.com. Pricing and limits change during early access — verify at the official site before committing budget.