Independent guide, not affiliated with or endorsed by TypeSafe AI. “Jev” is TypeSafe’s name for its model, used here only to identify it. Official documentation: docs.typesafe.ai.
Jev AI (TypeSafe): pricing, limits and independent measurements
Jev is a decision model, not a chatbot: you give it some text and typed questions, and it returns answers with probabilities in one pass. It costs $0.042 per million input tokens with output free. Below is what that means in practice, measured over 3,455 billed requests.
| Made by | TypeSafe AI, San Francisco (founded 2024) |
|---|---|
| Released | 15 September 2026; current version jev-1.13.0 |
| Price | $0.042 per 1M input tokens; output free; no subscription published |
| Measured per-request overhead | ~260 tokens on every call |
| Context | 64K per request (32K text + longest question); ~32K via gateways |
| Rate limits | 250,000 tokens/s and 1,200 requests/min |
| Measured latency | 375–734 ms warm (1K–32K tokens) |
| Direct signup | Paused since 22 September 2026; gateways still open |
What is Jev AI?
Jev is the first of what TypeSafe AI calls System One models, after the fast, intuitive thinking in Kahneman’s work. It was announced on 15 September 2026 alongside a US$40 million seed round led by DCVC. Instead of generating text token by token, it reads your input once and answers every question you ask about it in parallel, as one of three types:
- Choice returns the selected option, a probability for every option, and a confidence.
- Score returns a position on an ordered rubric (2–10 levels), per-level probabilities, and a confidence.
- Noul returns a yes/no probability between 0 and 1.
TypeSafe reports 40–200× faster and 40–400× cheaper than frontier LLMs on its own workloads; those are its figures, not independent ones. What we measured is below.
Is Jev AI free?
No ongoing free tier is published. Accounts opened while signup was open got $5 of credit, reported by VentureBeat, which buys about 119M input tokens: about 476K answered questions on a typical support-triage workload. Vercel’s launch offer of free Jev calls ended on 25 September 2026. What the credit covers for your workload.
How much does Jev AI cost per decision?
About $10.50 per million answered questions on a 500-token text with 4 questions per call. Two things decide the bill besides the price: every request bills about 260 tokens on top of your text, and the text is billed once per call however many questions you ask, so batching them is most of the saving.
At 100K items a day that is $126.00 a month. The cheapest LLM in our comparison, GPT-5.6 Luna, would bill $624.00, 5.0x Jev’s bill; GPT-6 Astra $29,700, 236x. Put in your own numbers or see every model ranked.
How do I get Jev AI API access?
TypeSafe paused new direct signups on 22 September 2026; existing accounts keep working. Without an account you can call the same model through OpenRouter, Vercel AI Gateway, Netlify AI Gateway, Cloudflare Workers AI, at the same rate where one is published, but with about 32K tokens of context rather than 64K. OpenRouter is the route we tested end to end. Every route, with a working call for each.
How fast is Jev AI?
TypeSafe states 70–500 ms end to end. Through OpenRouter we measured 375 ms for 1,000 tokens of text and 734 ms for 32,000, on a reused connection; the same small request took 1,630 ms when it had to open a connection first. Latency and rate limits in detail.
What are Jev AI’s limits?
250,000 tokens per second and 1,200 requests per minute, with a 429 past either; TypeSafe says both are still adjusting. A request can hold 64K tokens, and the text plus its longest question 32K; we measured the cut-offs at 65,536 and 32,768. A call with 1,000 questions still worked. Estimate your request’s size.
Can Jev AI chat or write text?
No. It returns labels, scores and probabilities, not prose. That makes it a replacement for an LLM wherever software needs a decision (routing a ticket, flagging a comment, checking an answer) and no help for writing, summarising or open questions.
Is Jev AI accurate?
We did not measure accuracy, and TypeSafe’s figures are its own. In one independent test with 128 candidate answers, Jev picked the right one 60% of the time and the open-weight Laya 39%. What we did measure is that its confidence is useful: across 960 paired answers we reworded or translated, the least confident third moved15.6 times as much as the most confident. Route low-confidence answers to a person or a fallback.
Jev AI vs GPT, Claude and Laya
On price per decision, Jev undercuts every LLM we track, from 5.0x against GPT-5.6 Luna to 236x against GPT-6 Astra, but it only answers typed questions. Jev vs GPT-5.6 Luna, all eight LLMs, and Jev vs Laya, the open-weight model you can run yourself, which answered a short question in 7.8 ms on our Mac.
Is Jev AI the same as JEV (Japanese encephalitis virus)?
No. This page is about Jev, the AI model from TypeSafe AI. JEV is also the usual abbreviation for Japanese encephalitis virus, an unrelated medical topic.
Get an email when direct signup reopens
TypeSafe gave no date. Leave an address and we will send one email when you can get a key direct again, or if the price changes.
Background from TypeSafe’s announcement and Wikipedia, read 2026-09-27. Price and limits checked against docs.typesafe.ai/models on 2026-09-28. Measurements: jev-fanout-bench, open code and data; the tables are also a Hugging Face dataset. Updated 2026-09-27.