Jev Launches Judgment-Only AI Model, Promises Faster Decisions at Lower Cost

A new AI model “Jev” developed by ChatGPT researcher Diogo Almeida has been released. It is distinguished by narrowing its functions to selection and judgment, without generating text. It is said to achieve 20-200 times the speed of conventional large models at 1/40 to 1/400 of the cost. Reporting by Qing Xiaowa of Xiaozhong Software details the development background and usage procedures.

He reportedly said of this approach:

To my knowledge, this is the shortest path to an AI-driven economic revolution

This article is based on Xiaozhong Software. The original author is Qing Xiaowa. The following summarizes the specifications and structure within the scope of publicly available information.

Design Philosophy Focused on Judgment Over

Generation

Jev is a judgment-focused model provided by TypeSafe. While conventional large models aim to generate fluent text, Jev concentrates computational resources on classification, scoring, and judgment. Output is limited to options, probabilities, and confidence levels. It does not generate long explanations or lists of reasons.

This trade-off is at the core of the design. In business automation, the reliability of conclusions matters most. The quality of prose is only secondary. Jev shortens responses by returning only judgment results.

The developer’s positioning also differs from conventional models. It is intended not as a replacement for general dialogue, but as a component in the decision-making layer. It is designed to be embedded in agents and business workflows.

How It Differs From Conventional Generative Models

The difference becomes clear with a concrete example. The task introduced in the summary is selecting a screenshot tool that meets specific conditions. The conditions are free, Windows-compatible, fast to launch, capable of simple annotations, and with no need for complex image management. The four candidates are ShareX, Snipaste, Flameshot, and PicPick.

When asked to a conventional model, a lengthy comparison and explanation is returned. The features, strengths, and weaknesses of each product are listed. Users must read through a large amount of text to reach a conclusion.

Jev’s response is limited to presenting the conclusion and probabilities. As an example, it is shown as Snipaste at 72%, ShareX at 19%, Flameshot at 7%, and PicPick at 2%. No extra explanation is included. Users can make an immediate selection by looking at the numbers.

Three Judgment Modes and How Probability

Output Works

Jev offers three main judgment modes. The first is Choice, which selects the best solution from multiple candidates. It is used for tasks such as selecting a deployment target, choosing a model to call, determining the next operation to execute, and identifying a submit button on screen.

The second is Score, which quantifies degree numerically. It evaluates in grades rather than simple yes or no. It suits measuring factors such as user dissatisfaction, content risk level, and fit to requirements. A graded evaluation from 0 to 3 is given as an example.

The third is a mode labeled Noul, which returns the probability of a matter being true or false. It is used to determine whether content is an ad, whether it is a refund request, or whether there is a safety risk. All modes output probabilities and confidence levels. Confidence is an indicator of how reliable the judgment itself is. In the example, a confidence level of 0.99 was shown.

Structure and Pricing Behind High Speed

and Low Cost

The reason for the speedup lies in structural simplicity. Conventional models generate text one token at a time. The longer the response, the more tokens are generated, increasing wait time and cost. Explaining reasons and creating comparison tables drive up the load.

Jev compresses a task directly into a judgment result. Processing ends once the result is produced. Because no additional explanatory text is generated, output tokens are kept to a minimum. The essence is that lengthy generation was excluded from the design stage.

Official pricing has also been disclosed. Input is said to cost $0.042 per 1 million tokens, with output free. $5 in credit is granted upon registration. For applicable tasks, speed is said to reach 20-200 times and cost 1/40 to 1/400. Note that these figures are the developer’s claims and are task-dependent.

Using Probabilities and Confidence to Power

Automation

The advantage of probability output is that it connects with operational rules. Unlike methods that simply return a name, processing can be branched based on likelihood. Real-world judgments often have no absolute correct answer. Having numbers makes it possible to set thresholds for automation.

The operational example introduced has three stages: over 90% is executed automatically, 60% to 90% requires additional confirmation, and under 60% is handed over to a human. By also using confidence levels, the stability of judgments can be factored in. This enables a division of labor where reliable processes are automated and uncertain ones are handled by humans.

This characteristic suits triage of routine tasks. It is effective for inquiry classification, risk assessment, and routing. In areas where misjudgment is costly, thresholds can be set strictly. Conversely, for minor tasks, the automation rate can be raised.

Setup Steps and Key Points for Getting Started

The steps to get started are simple. Register on the waitlist on the official website and enter your email address. According to reports, the wait for approval is short, completed in a few hours. After approval, register as a user via the sign-up link in the notice.

Issuing an API key follows the usual procedure. Select create on the dedicated screen, enter a name, and copy it. The key is shown only once, so it must be stored safely. It can be tried within the provided credit.

Skill installation is also provided. In Claude Code, run the marketplace-add and install commands. For other agents, use the skill-add command via npx. Detailed instructions can be viewed in TypeSafe’s skills repository. Paste-to-use is offered for Codex, Claude Code, Pi, Hermes, OpenClaw, Trae, Qwen Code, Kimi CLI, and WorkBuddy.

A key point for use is presenting candidates in advance. Jev is not suited for creating answers from nothing. It excels when choosing from existing information and candidates. For title selection and feedback analysis, candidates and criteria must be specified.

In the title-selection example, one is chosen from four options. The criteria are ease of understanding for general readers and appeal, reflection of core features, and avoidance of exaggeration. In the feedback-analysis example, classification and dissatisfaction level are determined at once. Classification candidates are bugs, feature requests, usage inquiries, and payment issues. Dissatisfaction is rated on a four-point scale from 0 to 3.

Role in Agent Integration and Outlook

Jev’s role lies in the judgment layer of agents. Generation and dialogue are handled by existing models, while selection is handled by Jev. Integration is expected to advance in areas such as selecting the next model to call, branching operations, and identifying on-screen elements. Because responses are short, latency from repeated calls can be kept down.

This direction overlaps with surrounding trends. In Visa Integrates Payments with ChatGPT to Enable AI Agent Purchases, the focus was connecting dialogue to execution. The reliability of judgments determines whether payments or orders can proceed. A mechanism that returns probabilities and confidence fits well as a gate before execution.

Meanwhile, the accuracy of generated content remains a challenge. As Lawyer Fined $5,000 Over ChatGPT Hallucinated Testimony shows, the cost of incorrect output is high. In automating judgments, assigning responsibility for misjudgments becomes essential. Threshold design and record-keeping become operational requirements.

Debate is also growing over AI use in family and daily life. In Altman Criticized Over ChatGPT Parenting Advice as Boundaries for Family AI Questioned, the issue was where to draw the line on use cases. A narrowed design like Jev could help clarify the scope of application. It can be seen as making explicit what it will not do.

Editorial Opinion

On short-term impact. In the next 3-6 months, separation of judgment processing is expected to advance in agent development. Cases are likely to increase where work shifts from configurations that have the same model handle both generation and judgment to a specialized layer like Jev. Low input unit pricing and free output lower barriers to experimentation and will accelerate replacement of classification and scoring, in our assessment.

On the long-term view. In 1-3 years, workflow design premised on probabilities is expected to spread. The practice of automating above 90% and reviewing the middle range is likely to become standardized, clarifying criteria for handover to humans. Recording confidence levels could become established as audit material, potentially supporting accountability for automation.

Questions from the editorial team. While the speed and low cost of judgments are verifiable, is probability calibration sufficient. Could accuracy be skewed by domain. Which tasks reproduce the developer’s claimed multiples remains to be measured by third parties, in our view.

References

• “Jev 入门：ChatGPT 研究员做了个“不会聊天、不写代码”的 AI 模型，但速度快200倍、成本低400倍”, by 青小蛙 — 小众软件, 2026-09-20T11:56:41.000Z (ARR)
• Source URL: https://www.appinn.com/jev-getting-started/