--- title: "What is Jev AI? A decision model, not a chatbot" description: "Jev replaces generated prose with typed choices, scores, and probabilities. That removes parser failures, not wrong decisions. Here is how to test it." canonical: "https://innovate-blog.com/articles/what-is-jev-ai" last-updated: "2026-09-20" --- # What is Jev AI? A decision model, not a chatbot > Jev replaces generated prose with typed choices, scores, and probabilities. That removes parser failures, not wrong decisions. Here is how to test it. By Moez Zhioua. Published 2026-09-20. Updated 2026-09-20. Category: AI models. Estimated reading time: 10 minutes. ## In brief - Jev returns bounded judgments for software instead of prose for a person. - A closed output schema prevents malformed answers, but Jev can still select the wrong valid answer. - Test one reversible decision on labeled cases before allowing a probability to trigger an action. ## The comparison with ChatGPT misses the useful part Five days after TypeSafe AI introduced Jev, search results were already full of explainers calling it a faster and cheaper alternative to a large language model. That comparison misses the useful part. Jev does not try to write what ChatGPT writes. It evaluates a piece of text or application state against questions whose possible answers were defined in advance. The response is a choice, a score, or the probability that a statement is true. Software can use that value without asking a text model to produce JSON and then checking whether the JSON survived. That narrower contract can remove an awkward layer from an automation. It does not make the underlying judgment correct. Any team considering Jev needs to test both halves of the system: whether the model makes the right call often enough, and whether the surrounding code knows when it should refuse to act. ## The job Jev refuses to do TypeSafe announced Jev on 15 September 2026 after two years of development. Founder Diogo Almeida previously worked at OpenAI on methods that contributed to ChatGPT. His new company starts from a criticism of the product he helped create: language models are trained to produce strings for people, while many software systems only need a bounded judgment. A support workflow may need to decide which queue owns a ticket, how urgent it is, and whether a refund is being requested. A chat model can answer those questions, but its natural product is text. The application has to instruct it to follow a schema, parse the response, validate the fields, and retry when the result does not fit. Jev removes the writing step. TypeSafe calls it a System One model, borrowing the name from fast human judgment. The company says it trained the model with Reinforcement Learning for Calibrated Decisions, or RLCD, and built it to return probabilities rather than prose. The result behaves more like a probabilistic function inside an application than a chatbot beside it. ## State goes in; bounded judgments come out A Jev request has two important parts. The state contains the information being evaluated. The questions define what the application wants to know and which answer shapes it will accept. TypeSafe currently documents three question types: - Choice selects one option from a set supplied by the developer. A ticket router might offer account access, billing, technical support, and other. - Score places the state along an ordered rubric. An incident could be rated from routine to severe, with each level described in plain language. - Noul evaluates a yes-or-no statement and returns the probability of yes. An application could ask whether a message includes a clear deadline. ## The clean response still depends on a good question Choice and Score return distributions across their allowed answers plus a confidence value derived from that distribution. Noul returns a number between zero and one. Several questions can share the same state in one request, although the documentation says each is evaluated independently. The developer still owns the difficult design work. The list of choices can omit the right answer. A scoring rubric can mix several ideas that should have been measured separately. A yes-or-no question can hide an ambiguous definition. Jev will stay inside the schema even when the schema is poor. ## No hallucinations describes one failure, not every failure TypeSafe says Jev cannot hallucinate. There is a precise version of that claim which follows from the product design: if the allowed choices are billing, technical, and other, Jev cannot return legal or a paragraph that breaks the parser. Its answer is constrained to the declared type. That eliminates an output-format failure. It does not eliminate a decision error. The model can choose billing when the ticket belongs with technical support. It can assign a high severity to a routine issue. A developer can set a threshold that sends too many uncertain cases into automation. Correct types prevent malformed output; they do not prove that the selected value matches reality. There is also a third layer after the judgment. The application decides what the answer means. Routing a ticket to the wrong queue is recoverable. Approving a transfer, publishing a change, or blocking a user has a different cost. The same model confidence should not automatically authorize all four actions. TypeSafe's confidence documentation recommends conservative thresholds, testing on the team's own data, and higher gates for consequential actions. The probability is an input to a policy. It is not the policy. ## The launch numbers need their labels TypeSafe lists Jev at $0.042 per million input tokens, with output left unmetered because the model does not generate a text sequence. The company reports end-to-end latency between 70 and 500 milliseconds. In its workflow evaluations, it reported gains as high as 193.6 times in speed and 444.6 times in cost against language-model alternatives. These are vendor results, not neutral benchmarks. TypeSafe provides more caveats than the headlines usually retain. It says the largest gains are likely at the high end of real workloads, the evaluation workflows were created internally, and the short demonstration used input that made the sampling difference especially visible. Its zero-hallucination result measures schema matching and is described as non-empirical. Independent evidence is promising but young. TechCrunch reported early developer tests with large speed and cost reductions. One email-classification comparison found Gemini slightly more accurate while Jev cost far less. Vercel reported that nearly 13 percent of its paid AI Gateway teams used Jev during its first 24 hours, the fastest uptake of any model launch on that gateway. Adoption proves interest. A low token price proves a low token price. Neither establishes that Jev will classify a particular company's cases well enough for unattended use. ## Where a decision model earns its place Jev fits when a workflow has unstructured input, a bounded answer space, repeated volume, and an error that can be measured. Ticket routing, document triage, agent tool selection, draft checks, and risk scoring can have that shape. Deterministic code remains better when the rule is known exactly. Dates, account balances, permission checks, and arithmetic should not become model judgments merely because the model is inexpensive. Jev is also the wrong component when the output must be an explanation, a summary, code, or a conversation. ChatGPT, Claude, or another generative model still owns that work. A common architecture will use both. Jev can decide whether a request belongs in a known path, code can enforce the permissions and thresholds, and a language model can write the human-facing response after the route is settled. The pieces should remain separate enough that a weak classification cannot quietly gain the authority of the executor. That boundary matters most for AI agents. A fast tool-selection score can reduce cost inside a loop, but the executor still needs narrow permissions, approval rules, a record of what happened, and a way back. The article on what an AI agent needs before it can act (/articles/what-ai-agent-needs-before-it-can-act) covers those controls. ## Test one decision before rebuilding a workflow Start with a judgment that people already make and for which past outcomes exist. Do not begin with an entire agent. ## 1. Define the decision and the consequence Write the allowed answers, the action attached to each answer, and the cost of a false positive and false negative. If the team cannot agree on those definitions, the model is not the first problem. ## 2. Build a labeled set Collect representative cases from the actual workflow, including rare but expensive failures, missing information, ambiguous inputs, and examples that should go to a person. Keep a separate holdout set for the final check. ## 3. Compare a real baseline Measure Jev against the current rule, classifier, or language-model call on the same cases. Record accuracy where it is meaningful, but also precision, recall, false-positive rate, latency, and total cost. Google's classification guidance notes that changing a threshold trades false positives against false negatives, so choose the metric that matches the consequence. ## 4. Set thresholds from observed errors Create an automatic band, a review band, and a refusal or fallback band. Use stricter gates for actions that are expensive or hard to reverse. Check whether the reported probabilities match outcomes on your cases instead of assuming launch-wide calibration transfers to your domain. ## 5. Run in shadow mode Let Jev make decisions without controlling production. Compare its calls with what operators actually did, inspect disagreements, and record the model version and schema. Promote one low-consequence path only after the error rate and fallback behavior are acceptable. NIST's AI Risk Management Framework adds the surrounding controls: name the people responsible for oversight, document the deployment context, test whether the system behaves as intended, and keep a contingency for failure in a third-party service. Those steps are ordinary operational discipline. A cheaper model makes them more necessary because teams will be tempted to call it more often. ## The schema becomes part of the product Jev's interesting bet is that many language-model calls are classifications and gates wearing a chat interface. Moving those decisions into a typed model can make an application faster, less expensive, and easier to compose. The tradeoff is easy to overlook. Once a model returns clean values, its answer can look like ordinary program state. The uncertainty has not disappeared. It has moved into the choice list, the rubric, the probability, and the threshold that turns that probability into action. For a first experiment, pick one reversible decision with labeled history. If Jev beats the current baseline at the error rate that matters, keep it. If the team cannot define the answer space or measure the wrong calls, a typed response only makes an unclear decision easier to automate. ## Sources and further reading - [Introducing System One Models & Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev), TypeSafe AI - [Introduction to Jev and System One models](https://docs.typesafe.ai/introduction), TypeSafe AI documentation - [Confidence](https://docs.typesafe.ai/confidence), TypeSafe AI documentation - [TypeSafe AI's Jev now available on AI Gateway](https://vercel.com/changelog/typesafe-ai-jev-now-available-on-ai-gateway), Vercel - [Jev is the fastest-adopted model in AI Gateway history](https://vercel.com/blog/ai-gateway-jev-model-launch), Vercel - [A new kind of AI model from a ChatGPT inventor is thrilling developers](https://techcrunch.com/2026/09/18/a-new-kind-of-ai-model-from-a-chatgpt-inventor-is-thrilling-developers/), TechCrunch - [Classification: accuracy, recall, precision, and related metrics](https://developers.google.com/machine-learning/crash-course/classification/accuracy-precision-recall), Google for Developers - [AI Risk Management Framework Core](https://airc.nist.gov/airmf-resources/airmf/5-sec-core/), National Institute of Standards and Technology Canonical URL: https://innovate-blog.com/articles/what-is-jev-ai