193 times faster and 444 times cheaper—Jev, which only makes "choices," has become a viral sensation.

193 times faster and 444 times cheaper—Jev, which only makes "choices," has become a viral sensation.

Over the past weekend, a model called Jev quickly gained popularity in the developer community.

It can't chat like ChatGPT, it can't write long articles, and it won't generate code in a serious manner.

Its operation is more like a frequently asked multiple-choice question: Continue or stop? Pass or block? Which button to click? Who should this email be forwarded to?

It seems like his abilities were "cut off" quite a bit, so why did he become popular?

Because Jev is aiming at one thing: judgment, making judgments based on "zero illusion".

Jev was spearheaded by ChatGPT co-founder Diogo Almeida. He wrote on X: “After co-inventing ChatGPT, I kept asking myself: Why didn’t a superhuman chat model bring AGI?” After two years of stealthy development, he offered his answer— a completely new decision-making model, perhaps even more effective than a stronger chat.

According to test data published on the TypeSafe website, Jev is 193.6 times faster and 444.6 times cheaper than existing large language models for automating workflow tasks. Input costs only $0.042 per million tokens, and output tokens are free. After launch, demand exceeded expectations, causing temporary API unresponsiveness.

Currently, Jev is open to all users, with no waiting list required; registration grants a $5 allocation (approximately 120 million tokens).

One person, two opposite things

Diogo Almeida's resume has an inherent tension.

He previously worked at OpenAI on the development of RLHF (Reinforcement Learning Based on Human Feedback) and InstructGPT, which directly led to the creation of ChatGPT.

Then he left and began to reflect on this method.

In a speech, he directly raised the question: "Today's AI can already challenge advanced mathematics, so why are companies still hesitant to entrust customer service decisions to it?"

His answer was: "Today's AI is very good at assisting, but not so good at automating."

Assistance means that you sit in front of the screen, AI does the work for you, and you check it.

Automation means that no human is present at all; AI makes judgments and executes tasks in the background on its own.

Almeida pointed the finger at RLHF itself. In his speech, he said, "Why do today's large models always need someone in the loop? Because when we train it, we literally put a person into that loop."

RLHF taught the model one thing: what kind of answers do people prefer?

This is an advantage for chat products. But for automated systems, it's a problem.

What is Jev: A function that makes conditional statements.

The TypeSafe website defines Jev as: "Taking unstructured state as input, outputting typed probabilistic decisions."

In simpler terms: You give it a piece of information, and it tells you what the answer is, and how confident it is. And it doesn't create any illusions.

It doesn't generate text. It only does three things:

  • Noul : A yes/no judgment, returning a probability between 0 and 1, representing the likelihood that something is true.
  • Choice : Select one item from a predefined list, supporting up to 255 options, suitable for categorization and routing.
  • Score : A score based on a set scale, used to measure urgency, quality, or risk.

Each response comes with a complete probability distribution and confidence level. Strongly typed output, no need to write JSON hints, and no additional parser required.

Why is it fast? Why is it cheap?

The way a typical large language model generates answers is by outputting tokens one by one, with each token depending on the previous one.

Jev's sampling method is parallel. All outputs are generated simultaneously in a single query. This is the fundamental source of the speed difference.

According to data from the TypeSafe official website:

  • Jev end-to-end response time: 70 milliseconds to 500 milliseconds
  • Current cutting-edge models: 3 seconds to 329 seconds
  • Speed difference: 40x to 200x (System One type tasks)

Regarding price:

  • Jev entered: $0.042 per million tokens ($42 per billion tokens)
  • Current cutting-edge model input: $0.20 to $10 per million tokens
  • Output Token: Free

TypeSafe's official website states that the figures of 193.6x speed improvement and 444.6x cost reduction come from its automated workflow evaluation, and that "these figures are expected to be at the higher end of the actual benefit range." The evaluation was conducted by members of the company's model capabilities team and may have some degree of bias.

"Calibration" is the real core.

Beyond speed and price, another key word that TypeSafe emphasizes is "calibration".

Existing large language models, even when asked to provide confidence levels, are often overconfident and inconsistent. The TypeSafe website states: "If a model can complete a task 95% of the time, but doesn't tell you when it's in that 5%, it cannot automate that task."

To address this, TypeSafe developed a new training method called RLCD (Reinforcement Learning for Calibrated Decisions).

The goal is: if the model says that something has an 80% probability, then in the batch of things that it judges to have an 80% probability, there should actually be about an 80% probability that something will actually happen.

In automated systems, this directly determines how tasks are allocated:

  • Confidence level > 90%: Execute automatically directly.
  • Confidence level between 70% and 90%: Submit to a stronger, larger model for verification.
  • Confidence level < 70%: Delegate to human decision-making

Earendil CTO Armin Ronacher explained: “Ultimately, this approach shifts the responsibility for dealing with hallucinations slightly onto the user. The user needs to decide: if the probability of this happening is only 50%, then perhaps it can be ignored. But if the probability reaches 95%, then I can take advantage of it. ”

Parallel decision-making: Asking ten questions at a time

Another feature of Jev is that multiple questions in the same request share the same input, but are evaluated independently and in parallel.

The WeChat public account "Digital Life Kazik" conducted a test: They submitted a news article and were simultaneously asked – Is it related to AI? Is it an advertisement? Has it received funding? Which category should it be classified under? Is it worth pushing?

Jev completes all these checks in one go, maintaining the same speed.

Traditional large language models require generating answers for each question individually, or stuffing all questions into a single prompt word before parsing and outputting the results.

This parallelism feature is most advantageous in scenarios with multiple problems.

What are developers already doing with it?

After Jev was launched, a large number of real-world application cases quickly emerged in the developer community.

Information feed filtering : Developer Marcel Pociot integrated with Jev to create a browser plugin for filtering unwanted content on the X platform. The average processing time for each post is 380 milliseconds.

Ad analytics : Someone used Jev to analyze 724 real-time ads in 40 seconds and made a total of 8,724 judgments.

Browser proxy : The open-source project jev-ultrafast from Browser Use uses Jev as the decision-making layer for browser operations. Taking flight booking as an example, each round of the program asks multiple questions simultaneously: Click or enter next? Which number to click? These questions share the same webpage state and are answered separately.

Context compression : By integrating Jev with Claude Code, we can determine which contextual content is critical information and achieve near-instantaneous context compression.

Security classifier : Vercel software engineer Pranit Sharma stated that replacing OpenAI's ChatGPT Luna 5.6 with Jev to run the command security classifier resulted in a 5 to 18-fold increase in processing speed and a significant improvement in accuracy.

Email Classification : Bryo AI's CTO, Nikhil Mudholkar, tested Jev and Gemini's performance in classifying business emails. His results showed that Gemini was slightly more accurate, but cost 10 to 20 times more.

Marketing analytics : One team integrated Jev into its marketing analytics process, scanning the Meta ad library and comparing ad format lifecycles, speeding up the judgment process that originally required manual work by 30 times and reducing the cost to less than $3.

Game Demo : TypeSafe's website showcases a demo of Jev playing Doom, making approximately 10 decisions per second. Jev makes real-time action judgments based on the structured game state, costing about $7 per hour. In the Ably Pong demo, Jev made 47 action decisions in 12 seconds, while Gemini, Claude, and GPT only made two or three in the same amount of time.

"Zero illusions" does not equate to "always getting it right".

Jev's official website states "Zero Hallucinations".

This statement is eye-catching, but also easily misinterpreted.

The technical specification defines TypeSafe as follows: Jev's output type and structure are predefined, so the model will not produce type errors, will not suddenly deviate from the specified format, will not fabricate a non-existent field, or will arbitrarily play out outside the limited options.

For example, the program only allows it to select from three options: "Pass, Reject, Manual Review". It will not output a poem or include a Markdown description.

This addresses the structural reliability issue in automated systems.

However, this does not mean that the model will not make a mistake. TypeSafe's own FAQ also explicitly acknowledges that Jev "can still make mistakes."

The emergence of Jev precisely addressed a real problem in agent development.

An agent that can manipulate web pages, call tools, and process files often has to go through a lot of minor judgments before completing a task.

Which element should I click next? Is the result returned by this tool valid? Which part of the context is no longer needed? Is this task truly completed, or just appears to be completed?

These judgments are not difficult to make on their own, but once they are incorporated into high-frequency workflows, costs will accumulate rapidly.

There are at least three clear reasons why Jev has become so popular this time.

First, its positioning is unusual enough.

While the industry continues to pursue longer reasoning, stronger chat, and greater context, Jev abandons text generation altogether and focuses solely on selection, classification, scoring, and routing.

Second, it addresses the cost issue of agents.

If a system makes hundreds of thousands of small judgments every day, a speed difference of 20 to 200 times and a cost difference of 40 to 400 times, even if it only applies to some tasks, is enough to attract developers to test it.

Third, it incorporates "uncertainty" into its interface.

In the past, models could always provide the answers; what really gives automated systems a headache is that models often don't know when they might make a mistake.

Jev's answer is: not only return the conclusion, but also the degree of certainty about the conclusion.

Why is it called Jev? Why is it called System One?

TypeSafe named the model category that Jev belongs to "System One Models".

The name comes from Daniel Kahneman's distinction between "System 1" and "System 2" in "Thinking, Fast and Slow": the former is more about fast, intuitive judgment, while the latter is more about slow, deliberate reasoning.

Jev is clearly betting on the former.

TypeSafe's product positioning is not to replace all large models with Jev.

The official list of applicable tasks includes classification, routing, scoring, information extraction, workflow branching, content review, model output verification, and batch processing on big data.

Complex reasoning, long text generation, coding, and content creation are still areas where traditional large language models excel.

Under this division of labor, Jev is responsible for making a large number of low-latency decisions first; when it is not certain enough, the problem is handed over to a slower, more expensive, but more suitable large model for generation and inference, or to humans.

Financing and Openness

TypeSafe AI was co-founded by Diogo Almeida, Erik Gafni, and Sasha Sheng, and the company operated covertly for about two years.

It officially debuted on September 15, 2026, and simultaneously announced the completion of a $40 million seed round of financing, led by DCVC.

TypeSafe states that they are currently the only company using this approach. Ronacher predicts that as the practical value of this model is gradually proven, more companies will follow suit.

When asked if the company is a frontier lab, Almeida said: "The main products of frontier labs are either fear or hype. I hope our main product is wisdom. We are not the kind of lab that is single-mindedly focused on creating unlimited wealth, or making religious narratives, or trying to create gods in the data center."

Risk warning and disclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.