A new kind of AI model launched recently, and it is worth understanding — not because it replaces what you already have, but because it does something your current stack almost certainly does not.
Jev, from TypeSafe AI, does not generate text. It does not reason through problems or write paragraphs. It takes a question, a typed list of possible answers, and returns a decision: which option, how confident, with a median latency under half a second. Probabilities, not prose. A verdict, not a conversation.
TypeSafe AI was founded by Diogo Almeida, previously at OpenAI where he worked on reinforcement learning from human feedback — the technique behind the alignment of the most capable models in the world. The move from training frontier models to building a pure decision engine is a deliberate one. Jev is not a smaller version of GPT. It is a different machine entirely.
The Sorting Machine
There is a piece of equipment at every major postal facility that reads labels, checks dimensions, and routes each package to the correct bin with extraordinary speed and precision. It is purpose-built. It does one thing.
Nobody would use it to write a letter. Nobody would call that a limitation. The sorting machine's value is not breadth — it is the depth of its single capability, delivered at a scale and speed that a general-purpose system cannot match.
Jev is an AI sorting machine.
The categories it sorts into are not postal bins. They are the typed options you define: departments, skills, queues, model tiers, routing paths, whatever the system needs to choose between. The question is not “what should I write?” but “which of these applies?” And the output is not a paragraph explaining the answer. It is the answer, with a confidence score, at $0.042 per million input tokens — in our testing, three to five cents per thousand decisions.
What This Solves in the Real World
The routing problem — assign this input to the right handler, quickly and at scale — appears everywhere in an AI-enabled business.
Support ticket triage
Billing question, technical issue, or cancellation request? A frontier reasoning model costs orders of magnitude more per call and takes seconds, not milliseconds. When you are processing thousands of tickets an hour, the economics matter.
Insurance claim routing
First notice of loss, supplemental claim, coverage dispute, or fraud flag? High volume, low margin per decision, consequences for errors.
Loan application sorting
Pre-qualification, income verification, credit review, manual underwriter queue? A bounded judgment that runs hundreds of times a day.
Content moderation queues
Safe, borderline, requires human review? The decision is simple; the volume is not.
Email-to-department routing
Sales, support, billing, HR? Every business solves this somehow. Most solve it badly.
Model selection itself
For each incoming request, does this need frontier-tier reasoning, a capable mid-tier model, or a local open-source model for privacy and cost reasons? That is a routing decision. It runs on every call.
In each case, the question is not whether AI can handle it. The question is whether a general-purpose language model is the right tool for a decision that is typed, bounded, and high-volume. A well-tuned stack does not use the same instrument for every task.
Where Jev Fits in a Well-Tuned Stack
The frontier models — reasoning-capable, multimodal, long-context systems that draft strategy memos and analyze legal documents — belong in a stack. So do capable mid-tier models for the bulk of executive output. So do open-source and local models for work that should never leave the building. And so do purpose-built capabilities like Jev, deployed precisely when the situation calls for them.
The operator question for an AI-enabled business is not “which model is best?” It is: which model is right for this decision, at this point in the workflow, given what this data is allowed to touch, at what cost per call?
Jev earns a place in that stack for the typed, bounded, high-volume decisions where a decision engine outperforms a language model on every axis that matters: speed, cost, and calibration on confident decisions.
We secured access to Jev while it was in beta. We aligned it into our stack — hardened adapter, contract-compliant data handling, deployment gates — and made it available to the Abeba AI Agent Fleet for exactly those sorting functions. The model is in position.
What We Found When We Tested It
We also tested it, because that is what a certification-oriented shop does before anything touches production routing.
Three rounds. Approximately 1,600 live calls. A vendor-agnostic harness, synthetic fixture sets, adversarial perturbation passes, independent blind labeling on contested decisions, sha-preregistered holdouts for final attribution.
What Jev proved — and it is genuinely strong
On routable traffic — requests the system actually knows how to handle — the calibration was flawless. In round 2: 183 confident routed decisions, 183 correct (precision at confident: 1.000). In round 3: 164 of 164 routed decisions correct. It routes on meaning, not surface pattern — paraphrase accuracy 93.8% on a set that stumped a word-matcher at 3.75%. Every error it made on ambiguous requests came at low confidence, which is exactly the behavior a production system can work with.
Where it sits on our shelf, and why
The test also found a boundary condition. On requests that fall entirely outside the defined option set, Jev's confidence signal does not reliably distinguish “wrong skill” from “nothing applies.” Roughly one in seven genuinely out-of-scope requests received a confident route to a real handler — the failure mode that executes the wrong tool with no tripwire. That is why Jev does not currently hold our production routing seat.
The deployable architecture is visible in the data: confidence-gated routing behind an external out-of-scope guard could close that gap. That is a different round, and a business case that would need to justify the added complexity.
For now: Jev sits ready in the stack for purpose-fit situations. It aces what it is built for. We know exactly where the boundary is, and we know that because we tested it before deployment, not after.
The sorting machine is excellent at sorting.
Deploying it well means knowing exactly which problem it solves — and which ones it does not.
That is what model selection looks like in a well-tuned stack.
Michael Murray
Michael Murray is the Managing Partner of Abeba Co, an AI accelerator that helps organizations build and operate intelligent agent systems. For more on building well-tuned AI stacks that compound with every interaction, visit abeba.co.