What Is Jev? Why Is Everyone Crazy About It? What I Built for Less Than $1.

LLMs talks through a problem. Jev makes a small, structured decisions. You can keep using Claude Code, Codex, or Perplexity Computer and add Jev to the workflows you build with them.

Karo (Product with Attitude)

Sep 21, 2026

TL;DR: On September 15, 2026, TypeSafe launched Jev, the first System One judgment model, and it became the fastest-adopted model in Vercel AI Gateway history: nearly 13% of paid teams used it within 24 hours. Pricing is $0.042 per million input tokens, with output tokens free. Measured on my own account: $0.000126 per four-question scoring request, so a $5 sign-up credit covers roughly 39,000 requests. The non-technical version: Jev is not another chatbot. It is a specialist at a sorting station. You hand it information and a fixed menu of options, and it hands back a decision your software can use directly. I built a newsletter SEO, AIO, and GEO dashboard with it in a few minutes.

Animated GIF showing three separate interfaces built with or using Jev. The first interface is a visual dashboard with category labels, image thumbnails, structured content cards, and metadata fields. The second is a scoring and evaluation interface with multiple criteria, numerical results, and horizontal status indicators used to assess content or structured inputs. The third is a conversational-style interface with a text input area and reaction controls. Together, the three examples demonstrate different ways Jev can be embedded into products: visual classification, structured scoring and evaluation, and lightweight user-facing decision interfaces. — Three different Jev interfaces: a visual classification dashboard, a structured scoring and evaluation tool, and a conversational-style decision interface.

Up until September 15, 2026, the models at your AI table belonged to one family: LLMs.

They came from famous houses such as OpenAI and Anthropic, and their family tree stretched to labs around the world.

They wrote with you, researched with you, helped you think through unfamiliar problems. Built and tested software.

LLMs are the chatty branch of the AI family. You can talk to them about almost anything.

Now, there is a new family at the AI table: System One models.

They’re not LLMs.

Its first member is called Jev.

And Jev instantly became the fastest-adopted model in AI Gateway history.

Line chart from Vercel comparing the share of paid AI Gateway teams using newly launched AI models during their first 24 hours. Jev, the probabilistic decision model from TypeSafe AI, shows the fastest adoption, rising from 0% at launch to nearly 13% of paid teams after 24 hours. GPT-5.6 reaches roughly 6.5%, Gemini 3.8 Flash about 5.5%, DeepSeek-V4.1-Flash about 4%, GLM-5.3-Flash about 3.5%, and Claude Fable 5.1 about 1.5%. Jev passes the other models within the first 12 hours and widens its lead through hour 24. Vercel reports that Jev became the fastest-adopted model in AI Gateway history, reaching more than twice as many paid teams as any previous model launch within its first day. — Jev became the fastest-adopted model in Vercel AI Gateway history, reaching nearly 13% of paid teams within 24 hours—more than 2× the GPT-5.6 family and over 6× Fable 5.1. Source: Vercel.

Why Did Jev Break Adoption Records?

Two reasons: price and speed.

In its own evaluations, TypeSafe found Jev to be 194× faster and cost 445× lower than the LLM comparison.

Jev costs $0.042 per million input tokens.

Output tokens are free. The docs don’t say why, but Diogo Almeida does. In the Jev launch video, TypeSafe's founder explained: “Output tokens are free because they’re too cheap to meter.”

After testing Jev throughout the weekend, I still hadn’t used up the $5 credit I received at sign-up. I had a working SEO/AIO dashboard by then. Jev was doing the scoring inside it.

I also scored the draft of the article you’re reading right now against four rubric questions: citation likelihood, claim extractability, first-party evidence, and structure. Each request cost me $0.000126 (!).

The $5 sign-up credit covers roughly 39,000 of them.

I’ll be lucky if Product with Attitude ever makes it to edition 39,000.

Table titled “What one Jev judgment request costs,” updated September 2026, showing the author’s real API usage for scoring a 2,000-word article draft with four judgment questions using Jev by Typesafe. The table reports 2,994 input tokens, free output tokens, a total list-price cost of $0.000126 per request, equivalent to 0.0126 cents, and estimates that a $5 credit would cover about 39,000 similar requests. Source: author’s own API usage, September 2026. — Table showing the real cost of one Jev judgment request on a 2,000-word draft: 2,994 input tokens, free output tokens, and $0.000126 per request at list price. A $5 credit would cover roughly 39,000 similar requests. Source: author’s own API usage, September 2026.

Forward this to whoever pays your AI bills. Give them a reason to enjoy opening an email.

What Is a System One Model

System One is TypeSafe’s name for models built to make focused judgments inside software.

You supply the information and define the question and allowed answers. Jev returns a choice, score, or probability.

These models don’t chat; they decide. Instead of generating paragraphs, they return fast, structured results your software can act on immediately.

Pink presentation slide titled “We Took The Opposite Research Direction,” contrasting chat-focused LLMs with System One Models. The left column, “Not Chat,” describes reinforcement learning from human feedback (RLHF) as optimizing LLMs for human preferences and instruction following. It attributes mode dropping, overconfidence, and reliability problems to RLHF. The middle column, “A New Model,” introduces System One Models designed for use by machines, with a new architecture, sampler, and training algorithm called Reinforcement Learning for Calibrated Decisions (RLCD). A small encoded text block appears on the right. — TypeSafe’s models are built for software to use directly. The slide contrasts RLHF-trained chat models with System One Models trained using Reinforcement Learning for Calibrated Decisions (RLCD).

How to Think About System One Models vs. LLMs

In my experience, technical ideas become far clearer when they are attached to a mental model: something you can picture in your head and connect to what you already know.

And familiar analogies are a great tool for that.

You already know that Jev belongs to a different model family than ChatGPT or Claude.

These two families have different jobs.

Comparison table explaining the difference between Jev and Claude Code or a similar agentic AI tool using simple analogies. The first row compares Claude Code to “a colleague who is a developer” and explains that an agentic coding tool helps users think and work, builds workflows and software, runs commands, and checks results. The second row compares Jev to “a specialist at a sorting station” and explains that Jev evaluates information against defined questions and possible answers. The table highlights the core distinction between agentic AI tools and Jev: tools such as Claude Code perform multi-step work and execute actions, while Jev is designed for fast, structured evaluation and decision-making rather than open-ended generation or autonomous task execution. — Table: Jev and agentic AI tools such as Claude Code play different roles. Claude Code acts more like a developer colleague that can build, execute, and verify work, while Jev acts more like a specialist sorting station that evaluates information against predefined questions and answer options.

Claude Code is the developer on your team. You discuss and build something together. Jev has a narrower role: it works like a specialist at a sorting station.

For example, it can examine an email and choose from a fixed set of options:

customer question
partnership request
newsletter
invoice
spam

Jev can’t decide to return a sonnet or a recipe for lentil soup. The output must fit the shape we defined.

Jev vs. an LLM

Comparison table showing how Jev differs from ChatGPT, Claude, and other large language models across eight dimensions. The table explains that LLMs are built to generate text, can return many forms of output such as prose, code, JSON, refusals, or invented answers, and generate responses one token at a time. Jev is built to return predefined decisions, only produces values or structures allowed by the question, and evaluates requested outputs in parallel. The table also compares strengths: LLMs are suited to conversation, writing, research, coding, explanation, and longer reasoning, while Jev is suited to classification, scoring, ranking, routing, verification, and fast workflow decisions. It further contrasts direct software use, uncertainty reporting, control, and human oversight: LLM outputs often need validation and models retain broad freedom over responses, while Jev uses predefined formats, returns probabilities and confidence for Choice and Score, lets code define options and thresholds, and can route low-confidence or high-risk cases to humans. — Table: Jev vs. ChatGPT, Claude, and other LLMs across output type, generation method, best use cases, uncertainty, software integration, control, and human oversight. The key difference: LLMs generate open-ended responses, while Jev is designed for fast, structured decisions that software can use directly.

The 3 Decisions Jev Can Make

TypeSafe currently supports three question types: Choice, Score, and Noul.

What I Built with Jev

It took me four minutes, at about a hundredth of a cent per judgment, to build an SEO, AIO, and GEO optimization dashboard for my newsletter.

Insane.

Why: I’ve been sharing more and more about how I optimize my Substack for AI search. This tool could make that workflow much faster. And cheaper!

How it works: I started with the fields Substack asks me to fill in, such as the SEO title and meta description. I gave each field its own criteria and let Jev do the scoring.

Animated interface showing a Jev-powered content scoring workflow for SEO, AIO, GEO, and virality optimization. The tool uses the focus keyword “typed judgments” and evaluates multiple content fields, including a social subject line, SEO subject line, social title, SEO title, and SEO description. Example copy includes “Stop prompt-and-parse. Start typed judgments,” “Typed judgments: replace prompt-and-parse with Jev,” and “Typed Judgments with Jev: Replace Prompt-and-Parse.” Each field is scored separately for SEO search engines, AIO AI Overviews, GEO generative engines, and virality potential, while a package score summarizes overall performance. The interface explains that each pause sends a Jev request with per-field and package-level scoring questions, returning structured scores and probabilities rather than open-ended text. The example demonstrates how Jev can be used as a fast evaluation layer inside an AI content optimization workflow for headlines, metadata, search visibility, AI search citation potential, and social performance. — My Demo: Jev evaluating content with typed judgments across SEO, AIO, and GEO criteria. The workflow uses structured Jev scores to judge each field and the content package as a whole.

What Other Developers Built with Jev

Builders on X also wasted no time putting Jev to work. Here are some of their projects:

Matthew Berman built an app that analyzed 724 live ads from 37 brands in 40 seconds.
Demo: Jev turning live ads into structured category intelligence. The system reads each ad, runs multiple typed judgments in parallel, classifies hooks, formats, and awareness stages, and flags uncertain cases for human review.
HeyStefan built a visual classifier that lets you type a concept like “things you can wear” and watch Jev evaluate a large set of objects against it.
Demo: Jev evaluating a broad set of objects against the typed judgment “things you can wear,” showing how structured semantic classification can be used to filter or rank candidates without relying on free-form text generation.

Side note: Jev doesn’t support images or video. The demo above was built in a way where the app displays images while sending Jev descriptions or other text about them.

What You Can Build with Jev

Use Jev for every workflow that follows this pattern: "read this, decide which bucket it belongs in, then do X".

Jev’s speed could also improve real-time applications where waiting several seconds for a language model would ruin the experience.

Some examples for newsletter writers:

Classify reader replies into praise, question, lead, complaint, or unsubscribe risk.
Rank draft headlines by click potential
Flag paragraphs that are too vague
Score article ideas by novelty

For product people:

Detect high-risk churn signals from support messages
Classify support tickets by urgency
Score feature requests by customer value
Route bug reports to the right team automatically
Classify user feedback into usability or performance
Score onboarding sessions for likelihood of activation or drop-off.
Score PRDs for missing assumptions

You can find even more inspiration on Jevable, because yes, there’s already an aggregator for projects built with Jev.

How to Get Started with Jev

There’s little point in sharing a step-by-step setup guide. Claude Code, Codex, Perplexity Computer, they can all handle the onboarding for you.

Create a free TypeSafe account, copy the starter prompt, paste it into your AI tool, and you’re ready to build. You’ll also need an API key. The official quick start walks through this.

Screenshot of the TypeSafe Jev Quickstart page showing the agent setup instructions for installing the TypeSafe skill. The setup box explains that Claude Code users can install the skill with the claude plugin marketplace add typesafe-ai/skills and claude plugin install typesafe@typesafe-ai commands, while users of other AI coding agents can run npx skills add typesafe-ai/skills --skill typesafe-ai and select their agent. The page also links to the TypeSafe skill documentation on GitHub and includes a “Copy Agent Prompt” button for copying the full installation prompt into an AI agent. — TypeSafe Jev Quickstart screen with agent setup instructions for installing the TypeSafe skill in Claude Code or another supported AI agent.

Can Non-Technical People Use Jev?

Yes!

Non-technical people can absolutely use Jev.

Just remember the mindset shift from the beginning of this post: you don’t use it the way you use ChatGPT or Claude.

You don’t sit down and chat with Jev directly. It doesn't generate text; it returns judgments. Your conversations still happen with your regular AI tool; Jev becomes the fast decision-making layer underneath.

Install the skill, stay in the AI tool you already know, and ask:

Review my workflows, skills, and suggest where Jev could help me save time and money.

I recommend using a model with high reasoning power for this.

Then ask your AI to help you build a demo so you can test and improve it.

And if API keys or configuration sound intimidating, don’t panic. Ask your AI to walk you through those too.

Keep Reading

What Jev Means for AI Agent Security. The Security Gamble Behind “Jev Can’t Hallucinate” with ToxSec

I’m Karo Zieminski, AI PM and builder. I write Product with Attitude, an AI newsletter for tens of thousands of readers across 146 countries, helping them develop critical AI literacy the only way it sticks: through practice.

SUBSCRIBE FOR FREE

GET CITED BY AI ⏐ GET FEATURED⏐ CONNECT WITH EVERYONE ⏐ SWAP LINKS

YOUR RESOURCES

AI AGENTS ⏐ BUILD⏐ DESIGN⏐ CRITICAL AI LITERACY ⏐ CLAUDE ⏐ PERPLEXITY

Fabrice Talbot

Sep 21

Thank you for sharing, this one is a keeper!

I looked it up online and there's no MCP yet which means you have to build the code yourself and call their API. It's not a big deal, just worth noticing. The skill helps you build the code/integration.

I really like the cost model. The real work is to decide where it goes inside your implementation and how to delegate the work safely. I am usually not a big fan of multi-LLM designs exactly for this reason. If you choose one LLM for each scenario, you may end up with 20+ and it becomes a maintenance nightmare.

But honestly here this one is a GEM because of its specific purpose. The separation is clear and easy to implement.

4 replies by Karo (Product with Attitude) and others