What Is Jev AI? TypeSafe's System One Model Explained, With What We Measured
TL;DR: Jev (
jev-1.13.0in our tests) is TypeSafe AI’s “System One” decision model. You give it text plus structured state and a question; it answers with a typed value and a probability, never free text. It is designed for many small judgments (classify, score, route, verify) where cost, speed and calibrated confidence matter. On our own test of 150 held-out Cantonese forum posts it was 3 to 5 points behind Claude Haiku 5.5 but cheaper, and its confidence was well calibrated.
Most people meet Jev through a one-line description. This page is the plain-English version: what it returns, how it differs from a chat-style LLM, what TypeSafe says it is for, where its docs say it is weak, and what we actually saw when we ran it. The full measurements live in the independent Jev vs Claude Haiku benchmark; the setup is in the how-to guide.
What is Jev AI in plain English?
Jev is a model that makes judgments instead of writing prose. You send a request containing some state (the text and any fields that matter) and one or more questions. It returns a typed answer for each question, with a probability or confidence, so your code can branch on the result without parsing a paragraph.
TypeSafe calls this family “System One” models: fast, intuitive judgments that code can combine like programming primitives. Our measured facts for the version we tested:
- Model:
jev-1.13.0(pinned), as of 2026-10-12. - Price: $0.042 per million input tokens; output tokens are free (TypeSafe docs). See Jev pricing.
- Limits: 80 requests per second and 100K tokens per second per account (docs).
- Language: English is the primary training language; the docs say CJK is “handled but not equally well”.
What are Choice, Score and Noul questions?
Jev has three question types. The examples below are generic (a support-ticket style), not results from our test.
| Type | You get back | Tiny example |
|---|---|---|
| Choice | One option from a list you define, plus confidence | “Which team should handle this ticket?” with options billing, bug, other |
| Score | A level on a scale you define (we used 5 levels) | “How urgent is this ticket?” from 1 (not urgent) to 5 (critical) |
| Noul | A yes-probability between 0 and 1 | “Does this ticket mention a refund?” returns, say, a number near 0 or near 1 |
You can ask several questions in one request; they are evaluated on the same state. In our sentiment work we used a Score for sentiment, a Noul for relevance and a Choice for intent.
How is Jev different from a generative LLM?
| Jev (System One) | Generative LLM | |
|---|---|---|
| Output | Typed answer plus probability | Free text, often JSON you must parse |
| Confidence | Built in, and calibrated in our test | Usually self-reported or absent |
| Text generation | No | Yes |
| Price shape | Input tokens only; output free | Input and output tokens |
| Cost per post in our test | $0.000045 | $0.000106 (Claude Haiku 5.5, batched 15 per request) |
| Speed in our test | Median 0.36 s per post | Median about 15 s per 15-post batch |
| Instructions | Read literally | More tolerant of vague wording |
The numbers in the lower rows come from our run on 2026-10-12 and are specific to a three-question sentiment, relevance and intent task.
What is Jev designed for?
TypeSafe’s docs list these task categories:
- Classification (intent, topic, department)
- Detection (spam, fraud, urgency, jailbreaks)
- Scoring (severity, relevance, quality)
- Routing (tool use, escalation, model routing)
- Search, retrieval (RAG context) and ranking
- Verification (citation support, policy violations, tool-call errors)
- ML feature extraction
- Structured data extraction (selecting among candidates)
We tested one of these: classification and scoring of forum posts for social listening. We have no evidence on the others, so we do not rate them.
What are Jev’s known weak spots, and what did we observe?
TypeSafe’s docs (“Jev 1.13 jaggedness”) list the following. The right-hand column is only what we saw; where it is empty, we did not test it.
| Documented weak spot | What we observed |
|---|---|
| Literal reading of instructions | Yes. With our first intent definition, criticism of a government minister was labelled “complaint” until we wrote that criticism of politicians and the government is other. Intent accuracy rose from 82.7% to 85.3%. |
| Choice option-order bias towards the first option | Small. With the sentiment options reversed, 98.1% of answers were unchanged (105 posts). |
| Large state full of irrelevant detail | Not tested directly. Quoted replies as one blob hurt: separating the quote from the reply raised sentiment accuracy from 62.5% to 70.0% on HardwareZone. |
| Sarcasm | Not a listed item, but we saw misses, as did Claude Haiku. Adding the opening post as context fixed one sarcastic Cantonese reply. |
| Math, dates, multi-hop, adversarial content, contradictory criteria, text generation | Not tested. |
When should you pick Jev, and when an LLM?
| Your situation | Better fit |
|---|---|
| Many small, well-defined judgments at high volume | Jev |
| You want calibrated confidence to auto-accept most rows and review the rest | Jev |
| Latency matters (real-time agent, stream filter) | Jev |
| You need text: summaries, explanations, rewriting | LLM |
| You need translation | LLM (Jev cannot replace it) |
| You need multi-step reasoning | LLM |
| You want the top accuracy on mild tone and intent and cost is irrelevant | LLM (Haiku led by 3 to 5 points in our test) |
| You want near-LLM accuracy at lower cost | Both: Jev on confident rows, an LLM on the rest |
Jev vs LLM: what were the results in one table?
Main test: 150 held-out LIHKG (Cantonese) posts, final definitions. Small sample, one forum, roughly ±6 points of noise.
| Label | Jev | Claude Haiku 5.5 | Hybrid |
|---|---|---|---|
| Sentiment | 78.0% | 82.7% | 82.7% |
| Relevance | 92.0% | 95.3% | 95.3% |
| Intent | 85.3% | 90.0% | 91.3% |
For definitions, calibration, per-language pilot results and the caveats, read the full benchmark. For what the same labels cost at scale, see Jev pricing and cost per post. For a concrete use case, see Cantonese sentiment analysis with LLMs.
How do you try Jev?
From TypeSafe’s quickstart: use the playground at console.typesafe.ai (log in, paste text as state, add Noul, Choice and Score questions), or create an API key and call the API. There is a Python SDK and a JavaScript SDK. Access was early access with a waitlist at launch (2026-09-15); the current docs simply say to log in, so sign in and check the console rather than assuming instant access. See the TypeSafe docs for the request shape.
What are the limits of this page?
- Our results come from one task, one forum for the main test, and 150 held-out rows. The answer key was written by an AI labeller (Claude) before any model was called, and a native reader spot-checked 20 of 20 labels as correct. That reduces but does not remove bias towards the Claude family.
- Versions are
jev-1.13.0andclaude-haiku-5-5as of 2026-10-12. - This is independent analysis. Neither TypeSafe nor Anthropic endorses or sponsors it.
Where we use it
We run these labels as an optional “AI sentiment labels” switch on our LIHKG actor on Apify, at $0.002 per labelled row with the first 20 rows free per run on the free plan, effective 2026-10-27.
From least effort to most freedom
FAQ
What is Jev AI? Jev is a System One model from TypeSafe AI. It reads natural language and structured application state and returns typed judgments with probabilities. It does not generate text, so it is not a chatbot, writer or translator.
What is a System One model? TypeSafe uses the term for models like Jev that turn natural language and application state into typed judgments and probabilities that code can combine, instead of producing free-form text.
How is Jev different from an LLM like Claude? An LLM generates text that you then have to parse. Jev returns typed answers (a choice, a score or a yes-probability) with confidence. In our test on 150 held-out Cantonese forum posts, Claude Haiku 5.5 was 3 to 5 points more accurate, while Jev cost less per post (about $0.000045 versus $0.000106) and gave calibrated confidence.
What is Jev good at, and what is it bad at? TypeSafe’s docs list classification, detection, scoring, routing, search, retrieval, ranking, verification and extraction as target tasks. Documented weak spots include literal reading of instructions, counting and math, date comparison, multi-hop reasoning, noisy state, and text generation.
Can Jev replace translation or summarisation? No. In our project it could not replace translation, which needs a generative model. Jev is built for many small, well-defined judgments.
Does Jev support Cantonese? Usefully but not perfectly. TypeSafe’s docs say English is the primary training language and CJK is handled but not equally well. We measured 78.0% sentiment accuracy on 150 held-out LIHKG posts; see the benchmark for the full numbers.
Further reading
- Jev vs Claude Haiku: independent benchmark
- How to use Jev for sentiment and relevance
- Jev pricing: cost per post and per million
- Cantonese sentiment analysis with LLMs
Chad runs 30+ published Apify actors, including the LIHKG scraper whose posts were used here. Numbers are from a run on 2026-10-12; raw posts are not published.