A call has two parts: the state, which is the text to judge, and the questions about it. There are three question types, and one call can mix them. Jev answers every question in parallel against the same state, so 60 questions cost about the same as one. Choice asks which of these options, and returns the chosen option, a probability for each, and a confidence. Score asks where on this scale, and returns a level, a probability for each level, and a confidence. Noul asks whether a statement is true, and returns a probability between 0 and 1. Here is a support ticket with two questions: which team gets it, and whether it is urgent. The request is POST /v1/systemone. The state is: “Hi, I've been trying to connect my Stripe account for 3 days and it keeps failing. I'm losing sales. Please help ASAP.” The model is jev-latest. department is a Choice, “Which team should handle this”, with billing for payment or subscription issues, technical for bugs or integration problems, and sales for pricing or account questions. is_urgent is a Noul: “The message conveys urgency or time-sensitivity”. In the abridged response, department chooses billing, with probabilities billing 0.84, technical 0.159, sales 0.001, and confidence 0.596. is_urgent comes back as noul 0.999. This example is abridged from TypeSafe’s quickstart. Read the department answer closely. Jev picked billing with a probability of 0.84, but its confidence is only 0.596, because technical also received some weight. The probabilities show which option Jev picked. The confidence shows how sure it is, and that is the number your code uses: act on the confident answers automatically, and send the rest to a person or to a larger model. One detail catches people out: the question ID, such as is_urgent, never reaches the model. Write the whole question in the instructions.

The model is Jev 1.13 (jev-1.13.0). jev-latest points to it. The price is $0.042 per million input tokens. Output is free. Rate limits are 250,000 tokens per second and 1,200 requests per minute. Context is 64k tokens per request, and 32k for the state plus the longest question. Input is text only: a string, a JSON object, or a list of strings. A Choice holds at most 255 options. Source: TypeSafe's models page. The SDKs retry with backoff when a request goes over a limit. On one page: a model that decides, and does not write. A state plus Choice, Score, and Noul questions go into the decision layer, which returns a typed answer, probabilities, and a confidence in 70 to 500 ms, all questions in parallel. The ticket about Stripe failing for 3 days gets department billing at 0.84, confidence 0.596, and is_urgent 0.999. Cost per case in the workflow evals runs from Opus 5 at $0.1761 down to Jev at $0.0004. Jev's 67.8% matches Sonnet 5, in 0.4 s against 78.1 s. The name comes from the Jevons paradox, 1865.

The official quickstart walks through Choice, Score, and Noul with the same support-ticket example. TypeSafe's launch post is the model, the evals, and the caveats on those evals. Akshay Pachaar puts the difference with an LLM plainly: the point is not that Jev generates faster. Jev does not generate text at all. A traditional LLM receives context and produces an answer one token at a time. Even when the output is a small JSON object, every token depends on those generated before it. Jev receives the same context but evaluates predefined decisions directly. When those decisions are independent, it can evaluate all of them in parallel. An agent handling a failed deployment may need to determine whether the incident is urgent, which team should handle it, whether the proposed command is risky, and whether the task is complete. An LLM generates a response containing these answers sequentially. The application then parses and validates it. With Jev, you define the questions and the expected answer types up front. It evaluates them together and returns typed answers with probabilities. The three primitives: Choice selects from known options, such as engineering, billing, or sales. Score places the input on an ordered scale, such as low, medium, or high risk. Noul evaluates a yes-or-no condition and returns the probability that it is true. The probabilities matter as much as the selected answers. If engineering receives 91% and billing 9%, automatic routing may be reasonable. If the probabilities are 52% and 48%, the system can escalate, gather more context, or call a stronger model. Control stays inside ordinary software. Code owns the thresholds and the consequences. Jev supplies the semantic judgment that a normal if statement cannot derive from unstructured text. It works best when the possible answers are known, the decision depends on meaning, and a careful person could judge the input quickly. It is not designed for writing, summarization, code generation, arithmetic, or decisions that need several dependent reasoning steps. Independent questions can run in parallel, but decisions that depend on earlier results must remain sequential. Jev also cannot return an option outside the declared schema, but it can still select the wrong valid option. Type safety prevents malformed outputs, not incorrect judgments. The mental model: LLMs generate new language when the answer space is open. Jev evaluates known paths when the answer space is bounded. Nathan Flurry's hype-free note is the other sentence: Jev does not replace GPT or Claude. It is a really smart switch statement, as if the classifiers of 2016 had received 2026 levels of intelligence. A new kind of tool, in the sense of a rebranded one, that will make a lot of workloads fast, cheap, and accurate. It needs a predefined set of options, and it will tell you which one to take. It cannot write code, generate natural language, reason step by step or show its work, produce any output you did not define in advance, or pick from more than about 255 options in one shot. It can classify, route, score, and rank. It can give a confidence. It can pick the right branch, tool, model, or sub-agent. It can judge, verify, or guardrail an LLM's output. It can label a very large number of rows. The workflows he pictures look like: an LLM proposes options, Jev decides, code executes. He also sees it fitting with code mode and MCP. Calling this a path to AGI seems far fetched to him, but he does not want to discount the applications it will make possible.

JEV GUIDE · LEARN

Что такое Jev?

Q: What people build with it

One public directory counts 729 builds. Most of them use Jev in one of a few ways: choosing the next action in an agent, sorting a queue into buckets, scoring every item in a large set, or stopping a risky step unless the confidence is high. In each case the system makes the same decision many times, so the speed and cost of each one add up. That directory's own counts: agents and browsers 167, benchmarks and evals 137, open source 126, SDKs and integrations 122, games and real time 74, routing and model choice 52, social feeds 51, tools and apps 51, coding and code review 39, search 36, context and memory 33, security and abuse 33, documents and OCR 30, trading and markets 25, inbox and support 24, ads and marketing 24, sales and leads 15, robotics and devices 14, UI 14, SEO and GEO 10, ecommerce 7. Jev Guide keeps its own showcases and GitHub projects, filed by content type and application. The counts on those pages are this site's. A few examples, as their authors published them: Jev plays Doom. Diogo Almeida wrote about 10 calls a second, about $7 an hour. Flight search with Browser Use. Gregor Zunic wrote that finding flights took 7 seconds and cost $0.0039, with a new action space every step, DOM as the state, and a small LLM only as a fallback for typing. The published card rounds the cost to about $0.004. 1kpapers. Hassan used Jev to classify 1,018 AI research papers: $0.08 in total, 256 ms median end-to-end latency per paper. The pipeline summarises each paper with DeepSeek V4 Flash, sends the title, summary, and 24 possible topics to Jev, then draws the result at 1kpapers.com. The summaries cost $3.99. The classifications cost $0.08. Instant compaction for Claude. tamara's question was why compaction, in 2026, is still a summarization prompt. Jev can score every tool call and drop what is irrelevant, and make that instant.

Q: Watch the founder explain it

Codez shared a 36-minute technical talk by Diogo Almeida, dated 19 September 2026. The paraphrase in that post is that the next era is not the Claude Code or Codex era, which still keeps a human in the loop, and that Jev is what comes after for LLMs: on the order of 200 times faster, 400 times cheaper, no value outside the type, and no human in the loop. The same talk is about why RLHF is no longer the thing, and how modern LLMs will be built. That is the poster's paraphrase, not a measurement by this site.

Jev is the first System One model from TypeSafe AI, released on 15 September 2026. You send it text and typed questions about that text. It returns typed answers, a probability for each option, and a confidence score.

Updated 19 Sep 2026

How a call works

A call has two parts: the state, which is the text to judge, and the questions about it. There are three question types, and one call can mix them. Jev answers all the questions in parallel against the same state, so 60 questions cost about the same as one.

Type	Asks	Returns
Choice	Which of these options?	The chosen option, a probability for each, and a confidence
Score	Where on this scale?	A level, a probability for each level, and a confidence
Noul	Is this statement true?	A probability between 0 and 1

Here is a support ticket with two questions: which team gets it, and whether it is urgent.

POST /v1/systemone

{
  "state": "Hi, I've been trying to connect my Stripe account for 3 days and it keeps failing. I'm losing sales. Please help ASAP.",
  "model": "jev-latest",
  "questions": {
    "department": {
      "type": "choice",
      "instructions": "Which team should handle this",
      "criteria": {
        "billing": "Payment or subscription issues",
        "technical": "Bugs or integration problems",
        "sales": "Pricing or account questions"
      }
    },
    "is_urgent": {
      "type": "noul",
      "instructions": "The message conveys urgency or time-sensitivity"
    }
  }
}

Response, abridged

{
  "department": {
    "choice": "billing",
    "probabilities": { "billing": 0.84, "technical": 0.159, "sales": 0.001 },
    "confidence": 0.596
  },
  "is_urgent": { "noul": 0.999 }
}

Abridged from TypeSafe’s quickstart. Read the department answer closely. Jev picked billing with a probability of 0.84, but its confidence is only 0.596, because technical also got some weight. The probabilities show which option Jev picked. The confidence shows how sure it is, and that is the number your code uses: act on the confident answers automatically, and send the rest to a person or to a larger model.

One detail catches people out: the question ID, such as is_urgent, never reaches the model. Write the whole question in the instructions.

Speed, cost and accuracy

Input costs $0.042 per million tokens. Output is free. Answers take 70 to 500 milliseconds. In TypeSafe’s own workflow evals, Jev matches Sonnet 5 on accuracy (67.8%) at $0.0004 per case, against $0.1174 for Sonnet 5.

Averages across four workflows, in workflow mode.
Model	Mean accuracy	Cost per case	Time per case
Opus 5	73.1%	$0.1761	37.8 s
Terra	67.9%	$0.0304	10.1 s
Sonnet 5	67.8%	$0.1174	78.1 s
Jev	67.8%	$0.0004	0.4 s
Haiku 4.5	53.6%	$0.0195	12.5 s

Source: TypeSafe workflow evals. The reference answers come from two large reasoning models, not from people. So 67.8% means Jev agreed with those two models 67.8% of the time. On a time chart Jev’s 0.4 s is almost invisible. On a cost chart, across the same four workflows, it is the short bar.

Models and limits

Model: Jev 1.13 (jev-1.13.0). jev-latest points to it.
Price: $0.042 per million input tokens. Output is free.
Rate limits: 250,000 tokens per second, 1,200 requests per minute
Context: 64k tokens per request; 32k for the state plus the longest question
Input: Text only: a string, a JSON object, or a list of strings. A Choice holds at most 255 options.

Source: TypeSafe’s models page. The SDKs retry with backoff when a request goes over a limit.

What is Jev on one page: a model that decides, the three question types, the Stripe ticket example, eval costs, and the limits.

Why it is called Jev

Jev is named after the Jevons paradox. TypeSafe’s founder, Diogo Almeida, wrote that in his launch thread. In 1865, William Stanley Jevons argued that as steam engines used coal more efficiently, Britain would burn more coal, not less. Make a decision cheap enough, and software starts making far more of them.

“System One” comes from Daniel Kahneman’s Thinking, Fast and Slow: System 1 is fast and intuitive, System 2 slow and deliberate. Jev is built for the fast kind.

On 15 September 2026 he wrote that, after co-inventing ChatGPT, he kept asking why superhuman chat models had not led to AGI. Two years in stealth produced a training method, RLCD, and this model: about 20 to 200 times faster, 40 to 400 times cheaper, output tokens free, built for decisions rather than for talk.

What people build with it

One public directory counts 729 builds. Most use Jev in one of a few ways: choosing the next action in an agent, sorting a queue into buckets, scoring every item in a large set, or stopping a risky step unless the confidence is high. The same decision happens many times, so the speed and cost of each one add up. Jev Guide files its own showcases and GitHub projects separately.

Agents and browsers167
Benchmarks and evals137
Open source126
SDKs and integrations122
Games and real time74
Routing and model choice52
Social feeds51
Tools and apps51
Coding and code review39
Search36
Context and memory33
Security and abuse33
Documents and OCR30
Trading and markets25
Inbox and support24
Ads and marketing24
Sales and leads15
Robotics and devices14
UI14
SEO and GEO10
Ecommerce7

Diogo Almeida@CompleteSkeptic𝕏

Демо Doom в реальном времени на Jev

Игровое демо показывает решения в реальном времени. Диого говорит, что оно делает около десяти вызовов Jev в секунду; в посте есть видео 2:43.

𝕏 · Игры

924354,8 тыс.1,2 млн

Gregor Zunic@gregpr07𝕏

Browser Use + Jev: сверхбыстрый браузерный агент

Gregor Zunic представляет открытый браузерный агент, объединяющий Browser Use и Jev: поиск рейсов за 7 секунд и 0,0039 доллара, с новым пространством действий на каждом шаге, пространством состояний DOM и резервной малой LLM для ввода.

𝕏 · Агенты

2649388,8 тыс.3 млн

Hassan@nutlope𝕏

Классификация 1018 статей об ИИ с Jev за $0,08

Автор суммировал каждую статью с помощью DeepSeek V4 Flash, затем передал заголовок, аннотацию и 24 возможные темы в Jev для классификации. Обработано 1018 статей об ИИ при общих затратах $0,08 и медианной сквозной задержке 256 мс на статью.

𝕏 · Аналитика данных

1181822 тыс.163,8 тыс.

tamara@tamarajtran𝕏

Мгновенное сжатие контекста с Jev

Пользователь делится идеальным сценарием использования Jev: мгновенное сжатие контекста. Оценивая каждый вызов инструмента и отбрасывая нерелевантное, Jev позволяет сжимать контекст мгновенно, без промптов-суммаризаций.

𝕏 · Инструменты разработчика

40496710,6 тыс.3,6 млн

Where it does not fit

It does not write text. Drafts, summaries, and code still need an LLM.
You define the possible answers. A Choice holds at most 255 options.
Input is text only for now: strings, JSON objects, and lists of strings. No images, audio, or video.
“Cannot hallucinate” means the answer always has the right type. It can still be wrong. Test the confidence scores on your own data before you choose a threshold.

Watch the founder explain it

Codez shared a 36-minute talk by Diogo Almeida, dated 19 September 2026. The post’s paraphrase: the next era is not the Claude Code or Codex era, which still keeps a human in the loop. The talk puts Jev after that, on the order of 200 times faster and 400 times cheaper, with no value outside the type. The same talk is about why RLHF is no longer the thing. That wording is the poster’s, not a measurement by this site.

Start here

Quickstart: state, questions, typed answersChoice, Score, and Noul, in one support-ticket example.Introducing System One Models and JevTypeSafe’s launch: the model, the evals, and their caveats.

Akshay Pachaar’s plain version: Jev does not generate faster. It does not generate text at all. An LLM writes an answer one token at a time, and even a small JSON object is sequential. Jev evaluates predefined decisions directly, in parallel when they are independent. An agent on a failed deployment may ask whether it is urgent, which team owns it, whether the command is risky, and whether the task is done. With an LLM, the app parses that out of a paragraph. With Jev, you declare the questions. Choice picks engineering, billing, or sales. Score places risk on low, medium, or high. Noul returns the probability that a yes-or-no statement is true. 91% against 9% can route automatically. 52% against 48% should escalate. Code owns the thresholds. Jev is the judgment an if statement cannot pull out of unstructured text. It is not for writing, summaries, code, arithmetic, or a chain of dependent steps. It cannot return an option outside the schema, and it can still pick the wrong valid one.

Nathan Flurry’s hype-free line: Jev does not replace GPT or Claude. It is a very smart switch. It needs a predefined set of options and tells you which one to take. It cannot write code, generate language, show its work, emit an output you did not define, or pick from more than about 255 options at once. It can classify, route, score, rank, attach a confidence, choose a branch or a tool, and label a great many rows. The loop he pictures is: an LLM proposes options, Jev decides, code executes.

Common questions

What is a System One model?

A type of model built for fast decisions that other software uses directly. The name comes from System 1, the fast, intuitive thinking in Daniel Kahneman's Thinking, Fast and Slow. It does not write text, generate code, or explain its reasoning.

Can Jev hallucinate?

It cannot return a value outside the type you asked for, so you never get broken JSON or an invented field. The answer can still be wrong. The confidence score tells you how far to trust it.

What does Jev cost?

Input costs $0.042 per million tokens. Output is free. A response takes 70 to 500 milliseconds.

Does Jev replace an LLM?

No. It handles small decisions that repeat, such as classify, route, score, filter, or approve a tool call. A larger model still writes anything that needs writing.

How do I call it?

With one POST request to TypeSafe's API, or through the Python or JavaScript SDK. You can try a question in the TypeSafe Playground first, without writing code.

What is Jev Engineering?

A way to build agents that took off on X three days after launch: an LLM writes, Jev makes the small decisions, and code acts on them.