Quick answer: what is Jev AI?
Jev is a new AI model from TypeSafe AI, released in early access on September 15, 2026. It does not chat or write text. You send it data plus typed questions, and it returns structured decisions with probabilities and confidence scores in roughly 70 to 500 milliseconds. TypeSafe prices it at $0.042 per million input tokens, and output is free.
In this article
What Is Jev? TypeSafe AI's First System One Model
Jev is the first public model from TypeSafe AI, and it takes a different road from the chatbots most people know. Instead of answering in sentences, Jev acts as a decision engine for software. A program hands it a block of information, called the state, along with a list of questions. Jev evaluates every question at once and returns answers in a fixed, typed format that code can use immediately. There is no text to parse and no broken JSON to repair.
TypeSafe calls this new category a System One model. The name comes from Daniel Kahneman's book Thinking, Fast and Slow, where System 1 is the fast, intuitive way we make quick judgments. The company describes Jev as a function call with frontier-level intelligence behind it: unstructured state goes in, typed probabilistic decisions come out.
Press coverage from MarkTechPost describes Jev as transformer-based but not a large language model, because it never generates text token by token. TypeSafe has not disclosed the architecture in detail.
Who Is Behind Jev? Meet TypeSafe AI and Diogo Almeida
TypeSafe AI is a San Francisco AI lab that came out of stealth on September 15, 2026. Launch-week reporting says it raised a $40 million seed round led by DCVC. The founder and CEO is Diogo Almeida, a former OpenAI researcher who is credited as a co-inventor of the reinforcement learning from human feedback (RLHF) methods behind ChatGPT. Some outlets spell his first name "Diego"; the company's own launch post uses Diogo.
Almeida's argument is simple. Chat models have been superhuman at conversation for years, so why is there still so little real automation? Models tuned to please human raters make great assistants, but automation needs answers that are consistent and honest about uncertainty. Jev is TypeSafe's attempt to build exactly that.
The name is also a nod to the economist William Stanley Jevons. His observation was that more efficient use of a resource can increase total demand for it. TypeSafe is betting that every big drop in the cost of intelligence will unlock far more uses for it.
How Does Jev Work? State In, Typed Decisions Out
Every Jev request has two parts: the state and the questions. The state can be a plain string, a JSON object or a list of text snippets, and it must be text only, so images, audio and video need to be transcribed or captioned first. The questions come in three types, and the whole API is built around them.
- Choice
- Pick one option from a set of up to 255, such as routing a support ticket to billing, technical or sales.
- Score
- Place something on an ordered scale of 2 to 10 described levels. The result can land between levels, for example how frustrated a customer sounds.
- Noul
- A yes/no question answered as a probability from 0 to 1, such as whether a message sounds urgent.
Jev evaluates all the questions in parallel and in isolation against the same state. Adding another question barely changes the response time, and you only pay for the extra question tokens. Choice and Score answers also carry a confidence value from 0 to 1, which TypeSafe derives from the shape of the probability distribution. Your code can act automatically when confidence is high and hand the case to a human when it is low.
A minimal example
from typesafe_sdk import Choice, Noul, TypeSafeClient
client = TypeSafeClient()
response = client.system_one(
state="I was charged twice for order A-104. Please refund the duplicate.",
questions={
"team": Choice(
instructions="Which team should handle this?",
criteria={
"billing": "Payments and refunds",
"technical": "Bugs and integrations",
"other": "Anything else",
},
),
"urgent": Noul(instructions="The message sounds urgent"),
},
)
team = response.answers["team"]
print(team.choice, team.confidence)What is RLCD?
Jev is trained with Reinforcement Learning for Calibrated Decisions (RLCD). Traditional chat models are tuned for answers that human raters prefer. RLCD instead rewards probabilities that are honest, so that higher confidence really does mean higher accuracy. That calibration is what lets developers set different confidence thresholds for different actions, with a low bar for reading a balance and a high bar for moving money.
Jev Speed and Pricing: What TypeSafe Claims
The headline numbers are what got developers talking. TypeSafe reports end-to-end response times of 70 to 500 milliseconds and says that is 40x to 200x faster than frontier LLMs on tasks shaped like System One queries. On its four-workflow evaluation, the company's homepage cites gains of 193.6x in speed and 444.6x in cost, though TypeSafe itself says those figures sit at the high end of what to expect in the real world.
| Metric | Frontier LLMs (TypeSafe's figures) | Jev (TypeSafe's figures) |
|---|---|---|
| End-to-end response time | 3 to 329 seconds | 70 to 500 milliseconds |
| Input price per 1M tokens | $0.20 to $10 | $0.042 |
| Output price | About 5x the input price | Free |
| Structured output errors | Possible | None by design: output must match your schema |
Treat the numbers as claims, not verdicts
At launch, these results were designed and run by TypeSafe itself, and no independent reproduction had been published. The company also says it cannot yet prove its low pricing is sustainable. Its main benchmark measures how closely models agree with the averaged answers of two frontier models (GPT-6 Astra and Claude Fable 5.1), not accuracy against a ground truth. On that measure, developer write-ups report Jev scoring about 67.8%, roughly level with GPT-5.6 Terra. Test Jev on your own data before you rely on the headline figures.
Jev vs Traditional LLMs: Key Differences
| Feature | Traditional LLMs | Jev (System One) |
|---|---|---|
| Main job | Chat, writing, coding, open-ended reasoning | Fast decisions inside software |
| Output | Generated text that usually needs parsing | Typed values with probabilities |
| Generation style | One token at a time, in sequence | All answers in a single parallel pass |
| Training focus | Human preference and verifiable rewards | Calibrated, honest probabilities (RLCD) |
| Confidence | Often overconfident when asked | Returned with every Choice and Score answer |
| Best for | Human-in-the-loop work such as copilots | Routing, scoring, filtering, guardrails, real-time loops |
| Not for | Latency-critical automation | Writing text, code or summaries |
What Can You Build With Jev? Real-World Use Cases
TypeSafe describes Jev as a way to add smart if-statements to ordinary code. Wherever a hand-written rule is too brittle, such as "is this message angry?", Jev can make the judgment and let normal code do the rest. The most promising uses include:
- Support ticket triage: classify, score urgency and route thousands of tickets in one pass.
- Content moderation and guardrails: score, judge and verify prompts and outputs from other AI models, including jailbreak detection.
- Relevance filtering for RAG: filter retrieved passages cheaply before an expensive model sees them.
- Real-time apps: game bots, robotics and simulations where a response in a fraction of a second matters.
- Large-scale data labeling: tag huge datasets at a cost that was previously impractical.
- Agent routing: let Jev decide which requests need a large LLM and which can be handled by plain code.
TypeSafe's launch demos include a bot that plays Doom, which the engineer estimated costs about $7 per hour at 10 decisions per second, and a Wikipedia wikiracing challenge. Other launch-week coverage mentions a Minecraft bot and a simulated drone navigating obstacles. Independent developers also shared early projects, all self-reported: one classified 1,018 research papers for about eight cents, and a browser agent completed a Google Flights search in about seven seconds for under half a cent.
Jev Limitations and Honest Caveats
Jev is impressive on paper, but it is a specialist tool. Knowing where it falls short will save you time.
Where Jev shines
- Very low latency and cost for high-volume decisions
- Answers always match the schema you define
- Confidence scores you can build thresholds around
- Many questions per call with little extra delay
What to watch out for
- It cannot write text, code or summaries
- Text-only input, with a 64k token limit for state and questions
- It reads questions literally, so vague wording hurts
- Counting, date math and arithmetic belong in your own code
- Accuracy can drop when the state contains lots of irrelevant material
- Text from users placed in the state can sway its answers
One point deserves extra attention. TypeSafe says Jev "can't hallucinate," but that claim is narrower than it sounds. Jev cannot return an answer outside the format you define, which removes malformed output and invented text. It can still pick the wrong valid answer, and that is why the confidence score and your own validation matter. The 0% error figure TypeSafe publishes for format errors is guaranteed by design rather than measured.
Access is also limited for now. Jev runs as a hosted API in early access behind a waitlist, and TypeSafe has not published model weights, a parameter count or a self-hosting option. Rate limits may change as more GPU capacity comes online.
Is Jev Worth Trying? Verdict and How to Get Access
Jev is best understood as a new building block, not a ChatGPT rival. The pattern early developer guides recommend is a cascade: Jev classifies and routes requests cheaply, ordinary code handles the simple cases, and a large LLM is called only for the hard minority. If you run high-volume classification, moderation, triage or real-time decision loops, a pilot is worth your time. If you need writing, coding or open-ended reasoning, keep using a chat model.
How to get started
- Join the early access waitlist at typesafe.ai.
- Create an API key in the TypeSafe console.
- Install an SDK:
pip install typesafe-sdkfor Python 3.10+, ornpm install @typesafe-ai/sdkfor JavaScript. A LangChain integration is also available. - Send your state and questions to the single
/v1/systemoneendpoint, or use the SDK. - Pin the model version if you tune confidence thresholds, and test everything on your own traffic first.
This article is independent analysis and is not affiliated with TypeSafe AI. Specifications, pricing and access may change, so confirm details on the official website.
Jev AI FAQ: Your Questions Answered
What is Jev AI?
Jev is a System One model from TypeSafe AI, released in early access on September 15, 2026. Instead of generating text like a chatbot, it takes program data plus typed questions and returns structured decisions with probabilities and confidence scores that software can use directly.
Who created Jev?
Jev was created by TypeSafe AI, a San Francisco lab founded by Diogo Almeida, a former OpenAI researcher known for work on reinforcement learning from human feedback and the research behind ChatGPT. The company emerged from stealth with a reported $40 million seed round.
Is Jev a large language model like ChatGPT?
No. Jev is not designed to chat, write or generate text. Press coverage describes it as transformer-based, but it produces all answers in one parallel pass and returns typed values, so it fills a different role from chat models such as ChatGPT, Claude or Gemini.
How much does Jev cost?
TypeSafe lists Jev at $0.042 per million input tokens, or $42 per billion, with output tokens free. The company cites $0.20 to $10 per million input tokens for frontier LLMs. TypeSafe admits it cannot yet prove the pricing is sustainable, so check its website for current rates.
How fast is Jev?
TypeSafe reports end-to-end responses of about 70 to 500 milliseconds and claims 40x to 200x speedups over frontier LLMs on tasks shaped like System One queries. These numbers are self-reported and were measured from the company's own setup, so test latency from your own servers.
Can Jev hallucinate?
Jev cannot return an answer outside the format you define, so it avoids malformed outputs and invented text. It can still choose a wrong valid answer, so treat its confidence scores as a signal and validate important decisions.
What is a System One model?
A System One model is TypeSafe's name for AI built for fast, structured decisions rather than text generation. The term nods to Daniel Kahneman's idea of fast, intuitive System 1 thinking, and these models return typed answers with calibrated probabilities.
What is RLCD in Jev?
RLCD stands for Reinforcement Learning for Calibrated Decisions, the training method TypeSafe uses for Jev. Rather than rewarding answers that human raters prefer, it aims to make probabilities honest, so higher confidence should correspond to higher accuracy.
How can I get access to Jev?
Jev is in early access as a hosted API behind a waitlist. Join through typesafe.ai, create an API key in the TypeSafe console, then use the Python SDK (typesafe-sdk), the JavaScript SDK (@typesafe-ai/sdk) or the REST endpoint. Model weights have not been released.
Will Jev replace ChatGPT or other chatbots?
No. Jev complements chat models. A common pattern is to let Jev classify and route requests cheaply, handle simple cases in ordinary code, and send only the hard cases to a larger LLM.
What does the name Jev mean?
TypeSafe named Jev after the economist William Stanley Jevons, whose observation about efficiency and demand inspired the company's bet that cheaper intelligence will unlock many more uses.
Sources and Further Reading
- TypeSafe AI: Introducing System One Models and Jev (official announcement)
- The Register: TypeSafe AI debuts model for machines that plays Doom
- MarkTechPost: TypeSafe AI releases Jev
- DataCamp: System One models and Jev
- LangChain: Building a harness with Jev
- DEV Community: How to use Jev, a practical guide