A support ticket comes in: "I was charged twice for my subscription." You want to route it to the right team without a human reading it first.
You could send it to ChatGPT and ask for a category. That works. It takes 3 to 30 seconds, costs a few cents, and usually comes back with the right answer. But "usually" is doing a lot of work in that sentence. Sometimes it adds a preamble ("Sure! Here is the classification:"). Sometimes it picks a category you did not ask for. Sometimes it wraps the answer in markdown that breaks your parser.
Now imagine a different tool. You send the same ticket and a fixed list of categories: billing, technical, account. The tool reads the ticket and returns one answer: billing, with a probability of 0.97. No preamble. No markdown. No chance of inventing a seventh category. The whole thing takes 70 milliseconds and costs a fraction of a cent.
That tool is Jev, and it is the first example of something called a System One model. It launched on September 15, 2026, and within 24 hours it became the fastest-adopted model in Vercel AI Gateway history, with nearly 13% of paid teams using it by the end of day one. Whether that early momentum holds is an open question. The interest is there, though, and the category is worth understanding before the noise catches up to it.
This post explains what a System One model is, how it differs from the language models you are used to, and when a decision model is the right tool for the job.
What Is a System One Model?
A System One model is an AI model that returns structured decisions with probabilities instead of generating text.
You send it context (called "state") and a set of typed questions. It evaluates all the questions in parallel and returns answers in a format your code can use directly: a category label, a score on a spectrum, or a yes/no probability. No text generation. No parsing. No formatting surprises.
The name comes from Daniel Kahneman's book Thinking, Fast and Slow. Kahneman described two modes of human thought: System 1 is fast, automatic, and intuitive. System 2 is slow, deliberate, and effortful. TypeSafe AI, the company behind Jev, borrowed this framing. Their argument: most decisions inside software are System 1 judgments ("which bucket does this go in?", "is this urgent?", "does this pass the check?") and we have been renting a System 2 engine (a language model) to make them.
That is the core idea. Not a better chatbot. A different tool for a different job.
How Jev Differs From a Language Model
Jev and a language model like GPT-5.6 or Claude Opus share a foundation: both are trained on large amounts of data, both understand natural language input, both use transformer-based architectures. But they are optimized for completely different outputs.
Here is how TypeSafe AI's launch post and documentation describe the difference, both published September 15, 2026:
Output format. A language model generates text one token at a time. Jev returns typed values: a choice from a set you define, a score on a spectrum you specify, or a boolean probability. The output structure is fixed before the query runs, which means type errors are impossible by construction.
Speed. Language models take 3 to 329 seconds for a response, depending on the model and reasoning depth. Jev takes 70 to 500 milliseconds. That is 40 to 200 times faster for comparable levels of intelligence on decision tasks, according to TypeSafe's own evaluations. Independent benchmarks have not yet been published.
Cost. Jev charges $0.042 per million input tokens. Output is free because the model does not generate text tokens. Frontier language models charge $0.20 to $10 per million input tokens, and output is typically 5x the input price.
Confidence. Language models can be prompted for a confidence estimate, but they tend to be overconfident and inconsistent. Jev returns calibrated probabilities with every answer, trained through a method TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD). Calibration means higher confidence correlates with higher accuracy across groups of predictions. It does not guarantee any individual answer is correct.
Hallucination. Language models can hallucinate: produce fabricated facts, invent tool calls, or generate text that does not match the expected format. Jev cannot produce text outside your declared schema, so it cannot hallucinate in the way language models do. It can still make wrong decisions. The structure is safe. The judgment is not guaranteed.
These are TypeSafe's claims, backed by their own evaluations. The speed and cost numbers are straightforward to verify. The intelligence comparisons need third-party reproduction. The company raised $40M led by DCVC and was founded by Diogo Almeida, who previously worked on RLHF and InstructGPT at OpenAI. The model is in early access through Vercel AI Gateway and TypeSafe's own API.
The Three Primitives
Jev's API is three question types. This is not a limitation you work around. It is the design.
Choice. You give the model a set of options and it picks one. Example: "Which team should handle this ticket?" with options for billing, technical, and account. Returns the chosen option, a probability for each option, and a confidence score. Supports up to 255 options.
Score. You define a spectrum with 2 to 10 ordered levels, described in words. The model returns a position on that spectrum. Example: "How frustrated is this customer?" with levels from "calm" to "angry." Returns a score that can land between levels (1.035), plus probabilities and confidence.
Noul. A yes/no question that returns a single probability from 0 to 1. Example: "Does this message request a refund?" Returns 0.95. No separate confidence field because the probability itself is the model's belief.
You can send all three types in a single query. The model evaluates them in parallel, which means asking 10 questions takes roughly the same time as asking one. This inverts the usual instinct to make a cheap call first and a follow-up only if needed. You ask everything upfront and let your code decide what was relevant.
When a Decision Model Is the Right Tool
System One models are not a replacement for language models. They are a complement. Each is built for a different layer of work.
Use a language model when you need text generation. Writing emails, drafting blog posts, summarizing documents, generating code, holding a conversation. These tasks need the flexibility of strings and the ability to reason through open-ended problems.
Use a decision model when you need a fast, structured judgment that your software can act on directly. Routing support tickets, classifying content, scoring urgency, checking whether an agent's proposed action is safe, evaluating whether a workflow should continue or stop. These are tasks where the answer is constrained, latency matters, and the output needs to slot into code without parsing.
A practical pattern we have started using: a language model drafts a response, then a decision model checks whether the response meets our criteria before it goes out. The language model decides what an agent should do next. The decision model evaluates whether that decision is safe to execute. Two models, two roles, composed in a system.
This is the same logic behind how to verify AI output: different tools for different layers of trust. A language model is good at generating possibilities. A decision model is good at evaluating them.
What This Split Means for People Building With AI
If you are starting out with AI, you do not need a System One model. A language model handles the work you are doing, and the tools are simpler. The value of understanding this split is practical, not theoretical.
Consider a content moderation pipeline. A language model can read a user submission and write a summary of the issue. That is a generation task. But deciding whether the content violates your policy, scoring its severity, and routing it to the right reviewer are decision tasks. Today, you might use a language model for all of it. You could replace the decision layers with a System One model and get faster responses, lower costs, and outputs that fit directly into your code without parsing.
The pattern is: language model for the creative work, decision model for the judgment calls. You compose them together. The language model writes the summary. The decision model classifies the severity. The language model drafts the response. The decision model checks whether it is safe to send.
We wrote about this idea in our guide to AI agents: agents use language models as their reasoning engine, but the surrounding system matters as much as the model. System One models are part of that surrounding system. They are the fast judgment layer that sits alongside the slow reasoning layer.
What We Do Not Know Yet
This is a new category, and the story is moving. A few things to watch:
Adoption durability. Jev's first-day adoption on Vercel AI Gateway was unmatched among recent launches, according to Vercel's blog post published September 18, 2026. But one week of data does not prove staying power. The next question is whether teams keep using it after initial experiments.
Independent evaluation. TypeSafe's performance claims come from their own benchmarks. The speed and cost numbers are easy to verify. The intelligence comparisons (matching frontier LLMs on decision tasks) need third-party reproduction.
Competition. Open-source alternatives appeared within days of Jev's launch. The category will likely have multiple players soon, which is good for users but makes the landscape harder to navigate.
Limits. Jev accepts text input only. No images, audio, or video. Context is limited to 64k tokens for state and questions combined. It gives up text generation entirely, which means it cannot explain its reasoning. If you need to know why a decision was made, you need a different tool.
None of these are reasons to ignore the category. They are reasons to pay attention to where it goes.
FAQ
What is a System One model?
A System One model is an AI model that returns structured decisions with probabilities instead of generating text. You send it context and a set of typed questions, and it returns answers in a format your code can use directly. The name comes from Daniel Kahneman's concept of fast, intuitive thinking in Thinking, Fast and Slow.
Is Jev a replacement for ChatGPT?
No. Jev is built for structured decisions inside software: routing, classification, scoring, verification. ChatGPT and other language models are built for text generation: writing, summarizing, conversing, coding. They serve different purposes and are designed to complement each other, not replace each other.
How fast is Jev compared to a language model?
According to TypeSafe AI's published benchmarks from September 15, 2026, Jev responds in 70 to 500 milliseconds. Frontier language models take 3 to 329 seconds for comparable tasks. That is a 40x to 200x speed difference. The speed advantage comes from parallel sampling: Jev generates all answers in a single pass instead of producing text one token at a time.
Can Jev hallucinate?
Jev cannot produce text outside your declared schema, so it cannot hallucinate in the way language models do. The possible outputs are defined in advance through the three primitives (Choice, Score, Noul), and the model selects from those options. It can still make wrong decisions. Calibrated probabilities tell you how confident it is, but confidence is measured across groups of predictions, not guaranteed for any individual answer.
Should I use a System One model for my project?
If you are building software that needs fast, repeated, structured decisions (routing tickets, classifying content, scoring inputs, verifying outputs), a System One model is worth evaluating. If you need text generation, reasoning, or open-ended analysis, a language model is the right tool. Many production systems will use both: language models for generation, decision models for evaluation and routing.
Want to go further?
If you are new to AI and want to build a foundation before diving into specialized tools like this, start with our beginner's guide to AI literacy. It covers the five core skills that help you decide which tool to reach for and when.
If you want to try Jev directly, it is available in early access through Vercel AI Gateway and TypeSafe AI's console. Early access is waitlisted, so you may not get in immediately.
All of this is just our opinion and should be fact checked. The model is new, the claims are self-reported, and the category is evolving. We will update this post as more information becomes available.