Everything you need to know about Jev and System One Models

Most AI automation in digital platforms today runs on large language models, even when the task has nothing to do with language. A form submission needs routing, a page needs tagging, or a comment needs checking before it goes live. The usual approach is to send a prompt to an LLM, ask for JSON back, and write code to handle the times it comes back wrong.

It works, but it's slow, it gets expensive at volume, and it breaks in ways that are hard to predict. You're using a model built to write text as a decision engine.

On 15 September 2026, TypeSafe AI released Jev into early access. It's the first of what TypeSafe calls System One Models. It isn't an LLM and it doesn't generate text. Its job is to make decisions inside software, quickly and at low cost.

Where the name comes from

The term comes from Daniel Kahneman's work on two modes of thinking. System 1 is fast and automatic. System 2 is slow and deliberate. LLMs have been pushed steadily towards System 2, with reasoning models spending more compute and more steps on each answer.

Jev goes the other way. It's built for the quick, repeatable judgements that make up most automation work. Is this email a complaint or a sales enquiry? Which of 40 categories does this article belong to? Does this submission need a person to look at it? None of these need a paragraph of text. They need a reliable answer and an accurate sense of how sure the model is.

TypeSafe was founded by Diogo Almeida, formerly of OpenAI. The team spent around two years in stealth building their own stack instead of fine-tuning an existing model.

How Jev works

You give Jev an input and a typed question. The input can be unstructured content, like an email, a support ticket or a page of copy, or structured application state. Jev returns one of three answer types:

Choice: it picks one option from a list of up to 255 that you define.

Score: it rates the input against ordered levels, such as low, medium, high and critical.

Probability: it gives a value between 0 and 1 for a yes/no statement.

Every answer comes with a confidence score. TypeSafe trains Jev with a method it calls Reinforcement Learning for Calibrated Decisions (RLCD), which aims for confidence scores that match how often the model is actually right. With LLMs, the training rewards answers people prefer, which is a different goal.

Calibrated confidence is the most useful part of Jev. It lets you set clear rules: act automatically above one threshold, send the item to a person below it, and escalate anything in between. That's much easier to govern than a model that sounds equally sure every time.

Under the hood Jev is still a transformer. It uses a parallel sampler to produce the full answer in one pass, instead of generating it piece by piece. TypeSafe hasn't published the architecture, the parameter counts or a technical paper.

The numbers

TypeSafe quotes response times of 70 to 500 milliseconds. On its own workflow benchmarks, it says Jev is around 190 times faster and 440 times cheaper than a leading frontier LLM. Pricing is US$0.042 per million input tokens, and output is free.

These are vendor figures, and TypeSafe itself says they're at the high end of what you'll see in practice. Early independent testing is more measured. One practitioner found Jev's accuracy was close to a fine-tuned BERT classifier (97.1% vs 97.6%), with much less setup. The biggest gains will go to teams that currently call a full LLM just to get a category or a yes/no answer.

The hallucination claim

Jev is being marketed as a model that can't hallucinate. The accurate version is narrower. Jev can only return values that fit the schema you define, so it can't invent a category, return broken JSON or drift into free text. That removes a class of production errors that make LLM-based automation fragile.

It can still pick the wrong option. The difference is that it tells you how confident it is, and the answer is always in a shape your code can use. Treat it as a well-calibrated classifier.

Digital experience teams

DXPs are full of small decisions that are currently made by rules, by people, or not at all. These include which personalisation variant a visitor sees, which content block suits a segment, whether a page meets brand and accessibility standards, and how new content maps to a taxonomy.

At sub-second speeds and this price, those decisions can run as part of the page request or the publishing workflow, instead of as an overnight batch. On Sitecore and Optimizely, that means real-time content selection and automated governance checks at publish time. It also helps with large migrations, where thousands of legacy pages need classifying, mapping and prioritising before they move.

Marketing teams

Marketing operations run on classification. Leads get scored and routed, form submissions get sorted, reviews and social comments get checked for sentiment and risk, and assets get tagged for reporting. Many teams haven't used AI here because per-call LLM costs add up at volume and the results vary.

At Jev's pricing, you can classify every enquiry and every asset instead of a sample. The confidence scores also give you a clear line between what's automated and what still goes to a person. That's an easier conversation to have with brand and compliance teams.

Technical teams

For developers, the main benefit is predictability. The output always matches your schema, so there's no parsing logic, no retry loops for malformed responses and no defensive validation around every call. Latency is low enough to sit in a synchronous request path, and costs are easy to forecast.

It also reflects a broader change in how AI systems are being designed. Good architectures no longer route everything through one large model. They use fast decision models for routing, gating and classification, and keep LLMs for work that needs language or multi-step reasoning. In an agentic workflow, a model like Jev can decide which tool to call, check whether an output passes guardrails, or flag a task for escalation.

Limitations

Jev is early. It's a hosted API behind a waitlist, with no self-hosting and no open weights. That will rule it out for organisations with strict data residency requirements for now. The architecture isn't public, the benchmarks are TypeSafe's own, and long-term pricing is unproven. It can't generate text, and the speed gains shrink on reasoning-heavy tasks. If you already run a well-tuned classifier, the improvement may be small.

It still deserves a proper test on your own data.

Getting started

Start by auditing your current workflows. Look for places where an LLM is used only to make a decision, and for repetitive triage that people handle against clear criteria.

Pick one, such as enquiry routing, content tagging or a publishing guardrail, and run Jev alongside your current approach on real data. Measure accuracy, latency and cost. Most importantly, check whether its confidence scores match actual outcomes, because that's what decides how much you can safely automate.

Where this is heading

Jev won't be the only model of its kind for long. The AI stack is splitting into layers. Fast, cheap, calibrated models will handle the steady flow of small decisions, and larger models will handle language, content and complex reasoning. The teams that get the most out of AI over the next few years will be the ones that design for that split early and use the right model for each job.