This week TypeSafe AI, a San Francisco research lab, opened early access to Jev, which it calls the first ‘System One Model’. Jev does not chat and does not write prose. It reads the information it is given and returns a structured decision from a fixed set of options, along with how confident it is. That is a narrow tool, and the narrowness is the point.
The announcement
What TypeSafe has actually released
Jev is an AI model built for software to call, not for people to talk to. A developer sends it the current state of a task, such as an email, a form submission or a document, together with a question and the answers it is allowed to give. Jev returns one of those answers in the expected format, plus a probability showing how sure it is. TypeSafe describes it as ‘unstructured state in, typed probabilistic decisions out’.
The company says Jev performs as well as the large general-purpose models on these decision tasks while being around 190 times faster and 440 times cheaper. Launch coverage reports responses in roughly 70 to 500 milliseconds, and TypeSafe lists a price of $0.042 per million input tokens, with output not charged. The team is led by Diogo Almeida, who launch coverage credits with co-inventing the training approach behind ChatGPT. The company has reportedly raised $40 million, and access is currently through a waitlist.
The name borrows from the psychologist Daniel Kahneman’s idea of two modes of thinking. ‘System 1’ is fast and intuitive, ‘System 2’ is slow and deliberate. Chat-based models have been pushed steadily towards the deliberate end, with longer reasoning and longer answers. TypeSafe is deliberately building for the fast, routine end.
A different tool
Why a decision model is a different kind of tool
Look inside most practical AI workflows and the model is rarely writing a report. It is answering small questions many times a day: is this enquiry a complaint, a sales lead or a delivery query? Which supplier is this invoice from? Is this request complete enough to process? Does this message need a person? Today those questions are usually sent to the same chat models that write essays and code, and the workflow then has to check that the reply is in a usable shape.
That mismatch has costs. Chat models can reply in an unexpected format, invent an option that does not exist or take several seconds to answer, and each call is charged as though it were a piece of writing. Teams add retries, validation steps and fallbacks to cope. A model that can only ever return one of the permitted answers removes a class of failure rather than managing it.
It also changes the economics of small, frequent decisions. If sorting an inbox, tagging documents or checking form submissions becomes very cheap and very fast, tasks that were not worth automating at chat-model prices may become worth a contained test.
The useful part
A confidence score is what makes human review workable
For businesses, the most interesting feature is not the speed. It is that every decision arrives with a probability, and TypeSafe says it has trained Jev so that those probabilities are ‘calibrated’: when it says 90 per cent, it should be right about nine times in ten. Chat models are notoriously poor at this. They tend to sound equally sure whether they are right or wrong.
A trustworthy confidence score turns ‘keep a person in the loop’ from a principle into a setting. A workflow can let high-confidence routine items continue, send uncertain ones to a named person and record both. The threshold becomes a business decision that can be tightened or relaxed as evidence builds up, rather than a choice between checking everything and checking nothing.
That only works if the confidence is genuinely reliable on your own work, which is something to test rather than assume.
- Define the options Write down the handful of outcomes each decision can have, including ‘needs a person’.
- Set a threshold Decide the confidence below which an item goes to review, and who reviews it.
- Check the calibration On a sample of real items, confirm that high-confidence answers really are right more often.
- Record every decision Keep the input, the answer and the confidence so mistakes can be traced and the threshold adjusted.
Read carefully
Treat the headline numbers with care
The speed and cost figures come from TypeSafe’s own tests, and TypeSafe itself describes them as ‘the higher end of real world gains’. Its accuracy is measured largely by how closely Jev agrees with other leading AI models, not against answers checked by people, and the company notes that the test workflows were written by its own team. Early independent checks reported in the press support the direction of the claims but covered a narrow range of tasks.
The product is in early access with no named production customers. TypeSafe says openly that it cannot yet prove its pricing is sustainable rather than subsidised. Jev cannot write text, summarise or hold a conversation, and each decision is limited to 255 possible answers. For a UK business there are also the usual questions about any new overseas service: where data is processed, what is retained and whether the terms suit the information you would send it.
None of that makes the launch unimportant. It means the sensible response is interest and a small test when access is available, not rebuilding a workflow around a week-old product.
The bigger picture
What it signals for AI more broadly
For the past few years the direction of travel has been one very capable model for everything, from drafting emails to writing software. Jev is a serious, well-funded bet on the opposite: specialised models for specific jobs, combined in a workflow. A chat model might draft the customer reply; a decision model decides whether the draft should be sent automatically or checked first; ordinary rules move confirmed information into the right system.
If that pattern takes hold, the difficult part of automation moves further away from the model and further towards the process. When a routing decision costs a tiny fraction of a penny, the limiting factor is whether the business has clearly defined what the options are, who owns the exceptions and what a good outcome looks like. Those are the same questions we encourage every SME to answer before choosing any tool.
What to do now
What a small business should do this month
There is nothing you need to buy. The useful step is to look at any workflow you already run, or are considering, and identify the points where AI is being asked to choose rather than write. Make each one explicit: the input, the allowed answers, the route for uncertain cases and the person responsible. That improves the workflow today, whichever model you use.
A workflow designed this way is also easier to change. If decision models like Jev prove reliable, swapping one step to a faster and cheaper model becomes a contained test with a clear comparison, instead of a rebuild. If they do not, nothing has been lost, and the workflow is clearer and easier to review than it was before.