Share some interesting applications of Jev: an AI model that does not make answers, only makes judgments

In a double-blind, pre-registration evaluation conducted by PrimeLine, one detail is telling.

Absolute accuracy of Claude Opus 5 and Haiku 4.5 on two real tasksComparable to Jev, or even better initially. But when the evaluation allows the model to skip samples for which it is least confident,Jev outperformed on both tasks.

The reason is that the confidence it outputs iscalibrated, while general generative models tend to be poorly calibrated and overconfident in their judgments. In other words: When "knowing what you don't know" is more important than "answering hard", the value of Jev comes out.

How to get started

If you are already using a model aggregation platform, the access cost is almost zero:

OpenRouter: Jev has been launched in the official model catalog, model name typesafe/jev, change the model name in the code and you can switch
Vercel AI Gateway: Integration completed within 72 hours after release, distributed via Vercel global edge network, with low latency
Official SDK: TypeSafe provides an online console (Playground) and complete documentation, supporting HTTP API, Python SDK, and JavaScript SDK
Official skill pack: TypeSafe has open sourced the skills warehouse on GitHub to teach coding agents to "ask all ambiguities at once" instead of just asking one question at a time.

Price: $0.042 per million tokens for input, free for output. There is also a free quota for new accounts to try.

Say something calm

First,It is not a replacement for larger models. Jev is a specialized decision-making model. Don’t expect it to have open reasoning and content generation capabilities. Its value is to achieve the ultimate goal of making quick and reliable decisions within clearly defined possible outcomes.

Second,In many cases, it is necessary to distinguish "whose credit". For example, the Browser Use demonstration of searching for flights in 7 seconds is the result of the cooperation of the entire tool chain, and all capabilities cannot be attributed to Jev. The community-organized case warehouse itself also marks the boundaries of evidence—which principles are clear, which effects need to be verified, and which claims lack basis. This attitude is worth learning.

Third,Be wary of blind blowing. Jev hit the screen within a week of its release. The popularity includes real technical value and a lot of marketing elements. What you really need is to run your own tasks with it and see how much delay and money you save.

But one thing is certain:AI applications are not just about “generation”. Outside of this track, the long-neglected infrastructure of “judgment” is being added. Whether Jev is the final answer is hard to say, but it raises the right questions.

The big model is responsible for writing answers, and Jev is responsible for making decisions - each performs its own duties, which is a good structure.

Reference: TypeSafe AI official release notes, Jev_apps case library (137 cases / 11 types of uses), Wikipedia, DataCamp technical analysis

This is one of the directions with the highest engineering value at present.

The Browser Use team releases in Jev only 3 daysAfter that, the official integrated jev-ultrafast was released, and it was said that it got 2,700 GitHub stars in two days. It splits web page operations into two sets of choices (action + goal),No screenshots or multi-modal tokens required, each round trip only produces two decisions.

Founder Gregor Zunic shared an uncut demo of Jev-powered browser automation completing a one-way flight search on a real airline website,The whole process takes 7 seconds, the model calling cost is about $0.0039.

There are a bunch of similar ones: Stagehand browser control, Cua's jev-use, TipTour that uses CoreML + OCR to locate the button on Mac and let Jev select the click location, voice control browser... There are currently 13 cases in this category, which is the category with the most cases.

4. Be a "router": help select the appropriate model for the task

There is a real waste in the multi-model era: using expensive models for simple problems is a waste of money. Jev’s low latency and low cost make it very suitable forTraffic triage.

Claude Code Router: Like work order triage, Jev first determines the difficulty of the task, and the plug-in then selects the sub-agent model and the reasoning strength of the main session accordingly.
Codex mission difficulty routing: Estimate the logical complexity of coding tasks before execution, simple editing is lost to cheap models, and complex cross-file reconstruction activates stronger models - it is reportedReduce the total model cost by more than 60%
Eve conditional routing/request distribution: First determine what type of model is needed for the problem, and then hand it over to the appropriate model for processing.
Division of local coding agents: Assign simple tasks, complex problems, and long-term tasks to different AI assistants

LangChain wrote a special article "Building Safety Belts with Jev", pointing out that the classification and routing logic in the agent framework are traditionally hidden in various closed source implementations. A general low-cost classification model like Jev can allow this layer of decision-making logic to be reused in open source agents. Armin Ronacher, founder of Sentry, also believes that it is economically unreasonable to use large generative models to make routing decisions.

5. Safety and Engineering: Those Inconspicuous Dirty Works

This type of application is not flashy, but it is the most practical.

Vercel replaces Shell command security classifier: Engineer Pranit Sharma replaced the internal OpenAI-based shell command security classifier with Jev. It is reported5-18 times faster, and the false alarm rate is lower - the reduction of false alarms actually improves the actual accuracy of the system
Prompt injection detection: The community has open sourced jev-mcp (an extension based on the MCP protocol). You can use Jev to do cross-source fact checking, detect prompt injection attacks hidden in user prompts, and score the semantic relevance of retrieved content.
Long context compression: fast-jev-compaction uses Jev to quickly evaluate the value of each tool call, delete useless logs while retaining key states - the context is easy to expand when coding agents perform long tasks. This project has also been recognized by the Jev official team
PR pre-screening: Analyze the changed file list and automated test results in the CI pipeline, initially evaluate whether the PR is eligible for merge, and automatically mark high-risk PRs and transfer them manually

6. Business pipeline: emails, work orders, customer service

What these scenarios have in common is:Need to make numerous small decisions quickly and consistently, using large models is expensive and slow.

Business email intent classification: Nikhil Mudholkar, CTO of BryoAI, compared Jev and Gemini. Gemini’s absolute accuracy is slightly better, but10-20x higher cost per call. More importantly, Jev provides calibrated probability distributions, making it easier for downstream systems to set confidence thresholds
Work order allocation and automatic routing: Digest the letter into key information and give it to Jev, select the most suitable one from a group of business teams, replace the fragile handwritten regular rules
Repeated deductions and emergency support classification: Quickly identify high-risk complaints (such as repeated deductions) in financial e-commerce scenarios. When customers are emotional or request refunds, they will be labeled as urgent within milliseconds and dispatched with priority.

7. Wilder gameplay

Generate image pixel by pixel: Developer @anshuc turns image generation into a pixel-level classification problem - giving each pixel a set of color options and probabilities, the program gradually fills it with the highest probability color selected by Jev, and finally spells out a low-resolution image
Automatic on-chain transactions: Some Web3 developers connect Jev to decentralized exchanges, read real-time order depth data, and only output three actions: buy, sell, and hold. The community feels that this is one of the most eye-catching applications, but the risks are also obvious under real market conditions.
Select components for the web interface: In the json-render project, Jev selects components and arranges relationships, and the program assembles the results into a renderable interface - like building blocks.

8. A noteworthy evaluation result

In a double-blind, pre-registration evaluation conducted by PrimeLine, one detail is telling.

How to get started

If you are already using a model aggregation platform, the access cost is almost zero:

OpenRouter: Jev has been launched in the official model catalog, model name typesafe/jev, change the model name in the code and you can switch
Vercel AI Gateway: Integration completed within 72 hours after release, distributed via Vercel global edge network, with low latency
Official SDK: TypeSafe provides an online console (Playground) and complete documentation, supporting HTTP API, Python SDK, and JavaScript SDK
Official skill pack: TypeSafe has open sourced the skills warehouse on GitHub to teach coding agents to "ask all ambiguities at once" instead of just asking one question at a time.

Price: $0.042 per million tokens for input, free for output. There is also a free quota for new accounts to try.

Say something calm

The big model is responsible for writing answers, and Jev is responsible for making decisions - each performs its own duties, which is a good structure.

Reference: TypeSafe AI official release notes, Jev_apps case library (137 cases / 11 types of uses), Wikipedia, DataCamp technical analysis

This is one of the directions with the highest engineering value at present.

4. Be a "router": help select the appropriate model for the task

There is a real waste in the multi-model era: using expensive models for simple problems is a waste of money. Jev’s low latency and low cost make it very suitable forTraffic triage.

Claude Code Router: Like work order triage, Jev first determines the difficulty of the task, and the plug-in then selects the sub-agent model and the reasoning strength of the main session accordingly.
Codex mission difficulty routing: Estimate the logical complexity of coding tasks before execution, simple editing is lost to cheap models, and complex cross-file reconstruction activates stronger models - it is reportedReduce the total model cost by more than 60%
Eve conditional routing/request distribution: First determine what type of model is needed for the problem, and then hand it over to the appropriate model for processing.
Division of local coding agents: Assign simple tasks, complex problems, and long-term tasks to different AI assistants

5. Safety and Engineering: Those Inconspicuous Dirty Works

This type of application is not flashy, but it is the most practical.

Vercel replaces Shell command security classifier: Engineer Pranit Sharma replaced the internal OpenAI-based shell command security classifier with Jev. It is reported5-18 times faster, and the false alarm rate is lower - the reduction of false alarms actually improves the actual accuracy of the system
Prompt injection detection: The community has open sourced jev-mcp (an extension based on the MCP protocol). You can use Jev to do cross-source fact checking, detect prompt injection attacks hidden in user prompts, and score the semantic relevance of retrieved content.
Long context compression: fast-jev-compaction uses Jev to quickly evaluate the value of each tool call, delete useless logs while retaining key states - the context is easy to expand when coding agents perform long tasks. This project has also been recognized by the Jev official team
PR pre-screening: Analyze the changed file list and automated test results in the CI pipeline, initially evaluate whether the PR is eligible for merge, and automatically mark high-risk PRs and transfer them manually

6. Business pipeline: emails, work orders, customer service

What these scenarios have in common is:Need to make numerous small decisions quickly and consistently, using large models is expensive and slow.

Business email intent classification: Nikhil Mudholkar, CTO of BryoAI, compared Jev and Gemini. Gemini’s absolute accuracy is slightly better, but10-20x higher cost per call. More importantly, Jev provides calibrated probability distributions, making it easier for downstream systems to set confidence thresholds
Work order allocation and automatic routing: Digest the letter into key information and give it to Jev, select the most suitable one from a group of business teams, replace the fragile handwritten regular rules
Repeated deductions and emergency support classification: Quickly identify high-risk complaints (such as repeated deductions) in financial e-commerce scenarios. When customers are emotional or request refunds, they will be labeled as urgent within milliseconds and dispatched with priority.

7. Wilder gameplay

Generate image pixel by pixel: Developer @anshuc turns image generation into a pixel-level classification problem - giving each pixel a set of color options and probabilities, the program gradually fills it with the highest probability color selected by Jev, and finally spells out a low-resolution image
Automatic on-chain transactions: Some Web3 developers connect Jev to decentralized exchanges, read real-time order depth data, and only output three actions: buy, sell, and hold. The community feels that this is one of the most eye-catching applications, but the risks are also obvious under real market conditions.
Select components for the web interface: In the json-render project, Jev selects components and arranges relationships, and the program assembles the results into a renderable interface - like building blocks.

8. A noteworthy evaluation result

In a double-blind, pre-registration evaluation conducted by PrimeLine, one detail is telling.

How to get started

If you are already using a model aggregation platform, the access cost is almost zero:

OpenRouter: Jev has been launched in the official model catalog, model name typesafe/jev, change the model name in the code and you can switch
Vercel AI Gateway: Integration completed within 72 hours after release, distributed via Vercel global edge network, with low latency
Official SDK: TypeSafe provides an online console (Playground) and complete documentation, supporting HTTP API, Python SDK, and JavaScript SDK
Official skill pack: TypeSafe has open sourced the skills warehouse on GitHub to teach coding agents to "ask all ambiguities at once" instead of just asking one question at a time.

Price: $0.042 per million tokens for input, free for output. There is also a free quota for new accounts to try.

Say something calm

The big model is responsible for writing answers, and Jev is responsible for making decisions - each performs its own duties, which is a good structure.

Reference: TypeSafe AI official release notes, Jev_apps case library (137 cases / 11 types of uses), Wikipedia, DataCamp technical analysis

Recently, a different model has emerged in the AI circle - Jev. It does not compete with GPT and Claude for the job of "writing answers", but instead focuses on a neglected position:make judgment.

Give it a current situation, add a question or a set of options, and what it returns is the answer to a multiple-choice question, a score, or a yes/no probability - and then the program executes it. The release date is September 15, 2026 (limited early access). The company behind it is called TypeSafe AI. The founder, Diogo Almeida, is one of the co-inventors of ChatGPT. The seed round is US$40 million, led by DCVC.

The concept sounds abstract, but the community has played hundreds of cases within a week. Pick a few of the most interesting ones and after reading them you will understand what problems they solve.

First understand the difference between it and ordinary large models

The essence of traditional large models is generative: you ask it, and it writes down token by token. I am good at writing articles, reasoning logic, and typing code, but when faced with a simple multiple-choice question like "choose one of these options", it becomes a overkill - slow and expensive.

Jev goes the other way: it's lightweightClassification and scoring models. Each time you call it, you give it a piece of background information and a set of clearly defined options, and it directly outputs the probability distribution of these options and returns the one with the highest confidence. The delay is tens to hundreds of milliseconds, and the cost is only a fraction of that of mainstream large models.

TypeSafe officially named itSystem One model——Borrowing the concept of "System 1" (intuition, quick judgment) in psychology and corresponding to the large language model of "System 2" (slow, deep reasoning). Official data is that it is 40-200 times faster than the leading large model, and the input cost is only $0.042 per million tokens.Output free.

1. Use it to play games: from Doom to 50 Subway Surfers

This is the type of application where it's easiest to understand the value of Jev - games require making small decisions many times per second, which is exactly its strength.

Official demo: Play Doom in real time. Jev makes about 10 action decisions per second, controlling character movement and shooting.One hour of continuous operation costs about $7——This number surprised the engineers, indicating that high-frequency real-time decision-making is feasible
Control 50 Subway Surfers simultaneously. Developer @_MaxBlade built an automated environment to allow Jev to control 50 games at the same time. Each action of changing lanes, jumping, and sliding is a small decision. Reportedly completed all 50 innings,Total cost less than one cent
Open Source Super Mario Controller. The system made by developer @faadilhshaik reads the structured game state directly from the memory. Jev decides when to jump and when to accelerate. The project has been open source
Speedrunning Slay the Spire 2. Chinese developer paulwei uses it to play strategy card games. When using large models in the early days, there was a significant delay between the two cards; after switching to JevEach decision takes only 0.7 seconds, the next command can be sent before the previous animation is finished.
Wiki Racing. Another official demonstration: starting from a Wikipedia article, navigating to the target article only by relying on page hyperlinks. Each step may face hundreds or even thousands of candidate links, and Jev needs to pick the one that is semantically closest to the target - testing the ability to choose among a large number of options.

2. Physical simulation: let the rocket land on its own

I particularly like this case because it proves that Jev is capable of more than just playing games.

Developer @uttkarsh_42 connected Jev to the MuJoCo physics engine to control a rocket that lands vertically. Jev is responsible for deciding:When to ignite, how many engines to run, and when to switch to landing attitude. After 12 tests, the system successfully landed. One of the complete experiments took 245 calls and reported a cost of approximately $0.04.

There is also autonomous driving in 3D city simulation: developer @jpschroeder integrated Jev into the virtual city. When the vehicle is driving, it uses sensor data and road conditions to determine the driving direction, and the new status is fed back to the system after each action. The entire build is said to have taken less than an hour.

3. Browser automation: search for real flights in 7 seconds

This is one of the directions with the highest engineering value at present.

4. Be a "router": help select the appropriate model for the task

There is a real waste in the multi-model era: using expensive models for simple problems is a waste of money. Jev’s low latency and low cost make it very suitable forTraffic triage.

Claude Code Router: Like work order triage, Jev first determines the difficulty of the task, and the plug-in then selects the sub-agent model and the reasoning strength of the main session accordingly.
Codex mission difficulty routing: Estimate the logical complexity of coding tasks before execution, simple editing is lost to cheap models, and complex cross-file reconstruction activates stronger models - it is reportedReduce the total model cost by more than 60%
Eve conditional routing/request distribution: First determine what type of model is needed for the problem, and then hand it over to the appropriate model for processing.
Division of local coding agents: Assign simple tasks, complex problems, and long-term tasks to different AI assistants

5. Safety and Engineering: Those Inconspicuous Dirty Works

This type of application is not flashy, but it is the most practical.

Vercel replaces Shell command security classifier: Engineer Pranit Sharma replaced the internal OpenAI-based shell command security classifier with Jev. It is reported5-18 times faster, and the false alarm rate is lower - the reduction of false alarms actually improves the actual accuracy of the system
Prompt injection detection: The community has open sourced jev-mcp (an extension based on the MCP protocol). You can use Jev to do cross-source fact checking, detect prompt injection attacks hidden in user prompts, and score the semantic relevance of retrieved content.
Long context compression: fast-jev-compaction uses Jev to quickly evaluate the value of each tool call, delete useless logs while retaining key states - the context is easy to expand when coding agents perform long tasks. This project has also been recognized by the Jev official team
PR pre-screening: Analyze the changed file list and automated test results in the CI pipeline, initially evaluate whether the PR is eligible for merge, and automatically mark high-risk PRs and transfer them manually

6. Business pipeline: emails, work orders, customer service

What these scenarios have in common is:Need to make numerous small decisions quickly and consistently, using large models is expensive and slow.

Business email intent classification: Nikhil Mudholkar, CTO of BryoAI, compared Jev and Gemini. Gemini’s absolute accuracy is slightly better, but10-20x higher cost per call. More importantly, Jev provides calibrated probability distributions, making it easier for downstream systems to set confidence thresholds
Work order allocation and automatic routing: Digest the letter into key information and give it to Jev, select the most suitable one from a group of business teams, replace the fragile handwritten regular rules
Repeated deductions and emergency support classification: Quickly identify high-risk complaints (such as repeated deductions) in financial e-commerce scenarios. When customers are emotional or request refunds, they will be labeled as urgent within milliseconds and dispatched with priority.

7. Wilder gameplay

Generate image pixel by pixel: Developer @anshuc turns image generation into a pixel-level classification problem - giving each pixel a set of color options and probabilities, the program gradually fills it with the highest probability color selected by Jev, and finally spells out a low-resolution image
Automatic on-chain transactions: Some Web3 developers connect Jev to decentralized exchanges, read real-time order depth data, and only output three actions: buy, sell, and hold. The community feels that this is one of the most eye-catching applications, but the risks are also obvious under real market conditions.
Select components for the web interface: In the json-render project, Jev selects components and arranges relationships, and the program assembles the results into a renderable interface - like building blocks.

8. A noteworthy evaluation result

In a double-blind, pre-registration evaluation conducted by PrimeLine, one detail is telling.

How to get started

If you are already using a model aggregation platform, the access cost is almost zero:

OpenRouter: Jev has been launched in the official model catalog, model name typesafe/jev, change the model name in the code and you can switch
Vercel AI Gateway: Integration completed within 72 hours after release, distributed via Vercel global edge network, with low latency
Official SDK: TypeSafe provides an online console (Playground) and complete documentation, supporting HTTP API, Python SDK, and JavaScript SDK
Official skill pack: TypeSafe has open sourced the skills warehouse on GitHub to teach coding agents to "ask all ambiguities at once" instead of just asking one question at a time.

Price: $0.042 per million tokens for input, free for output. There is also a free quota for new accounts to try.

Say something calm

The big model is responsible for writing answers, and Jev is responsible for making decisions - each performs its own duties, which is a good structure.

Reference: TypeSafe AI official release notes, Jev_apps case library (137 cases / 11 types of uses), Wikipedia, DataCamp technical analysis