Jev

Jev is an artificial intelligence model that makes decisions instead of writing text. It was released on September 15, 2026, by TypeSafe AI, a startup in San Francisco. A program gives Jev some information together with a set of questions whose possible answers are fixed in advance, and Jev returns one permitted answer to each question with a probability attached.^[1:1][2:1] A support system, for example, can hand it a customer's message and ask whether the case belongs to billing, technical support or account management; Jev answers "billing", and the program can route the case at once, with nothing to read or check.^[1]

Chatbots such as ChatGPT are built on large language models, which write their answers one small piece of text at a time. When a program needs only a decision from such a model, it must wait for the text and then pick the answer out of it. Jev skips the writing: it cannot produce replies, code or explanations of its reasoning.^[1:2] TypeSafe calls it the first "System One model", after the psychologist Daniel Kahneman's name for fast, intuitive thinking.^[3:1] The company says Jev answers in 70 to 500 milliseconds and costs a small fraction of what language models charge for the same tasks; these figures come from its own tests.^[3:2]

Jev drew wide attention among software developers. Within weeks, OpenAI, Amazon and Cloudflare released similar products, which the press grouped under the name "decision models", and researchers published dozens of imitations.^{[4:1][4:2][5:1]} In October 2026 TypeSafe raised $870 million at a valuation of $7.5 billion.^[6:1]

How it works

Every request to Jev has two parts. The state holds everything the decision depends on: for a refund request, the customer's message, the relevant transactions and the refund policy. The questions say what to decide about that state, and each has one of three types.^[1:1]

A Noul answer is not a plain yes or no but a number from 0 to 1: the probability that the answer is yes. In TypeSafe's words, "the number is the answer and the certainty in one." A value near 0.5 means the model is unsure. Programs usually turn the number into a decision with two thresholds; TypeSafe suggests treating values above 0.7 as yes, values below 0.3 as no, and sending anything in between to a person.^[7:1]

A single request can ask many questions about the same state. A language model asked for a structured answer still writes it out piece by piece, every bracket and label included. Jev instead answers all the questions together, in one pass through the model, which is the main reason it is fast.^[8:1][3:3]

TypeSafe charges only for the tokens (the word fragments a model reads) that go into a request: $0.042 per million. The answers cost nothing.^[9:1] By default a request may contain up to 64,000 tokens, and the state plus any one question up to 32,000. Jev is not retrained on a customer's data; it is adapted to a task only through the state and the wording of the questions.^[9:2] At launch Jev read text only;^[5:1] version 1.14, released on October 14, 2026, added PNG and JPEG images and a fourth question type, Box. It does not accept audio or video.^[9:3]

Training and architecture

TypeSafe says it built a new model architecture, a sampler that produces answers in parallel, and a training method it calls Reinforcement Learning for Calibrated Decisions (RLCD). It has published no technical detail about any of them.^[3:3] A model is calibrated when its probabilities match how often it is right: of all the answers it gives with 80% confidence, about 80% should be correct. TypeSafe notes that this holds across many answers and does not make any single answer reliable.^[1:3]

According to TechCrunch, Jev is based on the transformer, the same basic design used by language models, and was trained only on synthetic data, meaning examples generated by computer rather than collected from people. TypeSafe's chief executive, Diogo Almeida, declined to describe the architecture in detail, and outside observers suspect Jev is built on an existing open-weight language model, one whose trained parameters anyone can download.^[2:1] An engineer, Archer Hume, sent Jev 10,000 queries to study its behaviour and concluded that it holds the broad knowledge of a language model but, instead of predicting text, reads its decision probabilities directly from the model's internal representations.^[10:1]

Performance claims

TypeSafe describes Jev as 40 to 200 times faster than leading language models on the kinds of questions it is built for. On its homepage it gives headline figures of 193.6 times faster and 444.6 times cheaper, taken from its own workflow tests, which it expects to be at the high end of real-world gains.^[3:2] It also says Jev "can't hallucinate", meaning it cannot invent an answer outside the options it was given. The company presents this as a guarantee that follows from the design rather than as a measured result.^[3:4]

TypeSafe lists several limits to its own tests. Its own staff wrote the test workflows. The "correct" answers were taken from an average of two large models by OpenAI and Anthropic, which may favour those companies' models. The competing language models were run through TypeSafe's own software, which forced them to give structured answers. Speeds were measured mostly from laptops on the US West Coast. TypeSafe also says it cannot prove that its prices are not subsidised.^[3:5] The Register noted in September 2026 that few public benchmarks of Jev existed.^[10:2] One independent report came from the cloud platform Vercel, whose engineer Pranit Sharma said a classifier that checks commands for safety ran five to 18 times faster, and more accurately, after Vercel moved it from an OpenAI model to Jev.^[2:2]

Name

"System One" comes from Kahneman's distinction between System 1, the fast and intuitive mode of thought, and System 2, the slow and deliberate one; TypeSafe casts language models as System 2 and Jev as System 1. "Jev" honours the 19th-century economist William Stanley Jevons, who observed that more efficient steam engines increased the total amount of coal burned, because coal became worth using for more things. TypeSafe expects cheaper decisions by machine to spread the same way.^[3:1][2]

History

TypeSafe was founded in 2024 by Diogo Almeida, Sasha Sheng, a former research engineer at Meta, and the engineer Erik Gafni.^[6:1] Almeida had been a researcher at OpenAI; according to TechCrunch, he helped build ChatGPT and helped invent reinforcement learning from human feedback, the training method that taught language models to follow instructions. He has argued that this work optimised models for human language, while automating work in software needs output made for machines.^[2:3]

Jev launched on September 15, 2026, in early access with a waitlist. TypeSafe aimed it at sorting, routing, scoring and checking steps inside software, at processing very large datasets, at applications that must respond in real time, and at supervising other AI systems, for instance by spotting attempts to trick a chatbot into breaking its rules.^[3:6] Demand in the first days briefly exceeded what TypeSafe could serve.^[2:4] Developers built demonstrations that played the video games Doom and Tetris and the game of chess; at chess it lost to an open-weight model but cost far less to run.^[10:2] Vercel, which resells AI models to developers through a single service, reported that 35% of its paying teams used Jev within two weeks, nearly twice the share of the next model and the fastest uptake Vercel had recorded. TechCrunch called the data "a necessarily limited view", since it covers one platform only.^[11:1]

On October 9, 2026, TypeSafe announced that it had raised $870 million at a valuation of $7.5 billion, in a round led by the venture firm Andreessen Horowitz with Sequoia and DCVC. The company said a third of Fortune 500 companies already used Jev; it has not named them.^[6:1]

Reception

The developer Sean Goedecke argues that Jev's main advantage can be had from ordinary language models: if the answer must be one of a few options, a model can be made to produce just one short piece of text, which he found two to three times faster with a small open model. He calls the claim that Jev cannot hallucinate "a semantic dodge", because a model that cannot invent an answer can still pick the wrong one.^[8:2] The YouTuber Mo Bitar, who reviews AI products, asked: "I know it's fast. I know it's cheap, but is it good?" The AI researcher Andrej Karpathy said Jev revealed demand for fast models of "acceptable intelligence" that the race for ever-smarter models had overlooked. The Register added that Jev asks more of developers up front than a chatbot does, because each task must be broken into specific questions with fixed answers.^[10:2]

Competing decision models

In the weeks after Jev's launch, researchers published dozens of similar models.^[4:2] Large companies followed. OpenAI announced a similar service at the end of September 2026. In the same week, Amazon Web Services released Strands Decider 2B, an open-source model built on the language model Qwen3.5-2B, which began as a personal project of the Amazon engineer Marc Brooker; Brooker said such models "make a perfect decider for a workflow step".^[4:1] Cloudflare released two models, Clef and the smaller Clef-flash, which accept the same requests as Jev and can also read images and video. Their parameters are free to download, but their training data is not public, and Cloudflare's claim that Clef outperforms Jev rests on its own tests.^[5:1]

Almeida responded that imitators might be underestimating "the difficulty of making the models actually smart", and said he did not yet see real competition.^[4:2]