Jev 101 – A new class of model

Typesafe AI released Jev, a discriminative "System 1" model that takes a State JSON object and typed Questions and returns calibrated probability distributions for choice, score, and true/false outputs in a single pass rather than generating tokens. The model is priced at $0.042 per million input tokens with free output and is positioned for fast, cheap decision-making in production pipelines, in contrast to slower System 2 LLMs such as Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens.

In this post, we take a look at Jev, a new “System 1” type of model optimized for extremely fast and cheap decision making with calibrated JSON outputs. In a future post, we will compare it with Claude Opus 5.5 and decide when to use one or the other. If you want to learn more about Jev, consider signing up to my Maven cohort, “ Jev vs Claude: Find the Cheapest Model That Meets Your Quality Bar https://maven.com/bruno-goncalves/jev-vs-claude-a-measured-guide ” taking place on November 18 Meanwhile, we hope you enjoy this post, so and to help us grow. In his ground breaking 2011 book, “ Thinking Fast and Slow https://link.amazon/B09Q0rXpx ”, Daniel Kahneman introduced the idea that your brain has two different modes of thinking, what he called System 1 and System 2 . System 1 is the auto-pilot: fast, automatic, instinctive , and effortless. It runs continuously in the background and makes judgement calls: he looks angry, the road is blocked, this sounds off, etc. System 2 is the opposite. Slow, deliberate, thoughtful, rational . It follows the rules, makes calculations, explains itself, and pays attention. Unfortunately, attention is limited and System 2 is lazy so it often follow the lead of System 1. The current generation of LLMs, like Claude, ChatGPT, Gemini, Grok, etc fall cleanly into the System 2 category. They work through the prompt systematically to figure out what to output. They excel at producing long form text, work through complex problems, translate across multiple languages, etc. They fall neatly into the System 2 category. Despite these strengths, many of the pipelines that leverage LLMs in production revolve around quick decision making. Which team should handle this ticket? Is this summary correct? What tool should be called? System 1 type problems. A few weeks ago, on September 15, 2026, Typesafe AI released Jev, the first System 1 AI model https://typesafe.ai/blog/introducing-system-one-models-and-jev . Jev is a discriminative model https://en.wikipedia.org/wiki/Discriminative model not a generative one that specializes at producing quick, accurate, structured answers. Instead of a prompt, it processes a State and a set of Questions and it produces a probability distribution over the different possible answers for each question. It’s incapable of stringing two words together but it can make decisions quickly and cheaply. It costs $0.042 per million input tokens, output is Free , and generated in a fraction of a second https://www.tomshardware.com/tech-industry/artificial-intelligence/typesafe-ais-jev-offers-an-alternative-to-llms-that-claims-to-be-193x-faster-and-445x-cheaper-system-one-type-model-is-bespoke-for-probabilistic-decision-making . Claude Opus 5.5 https://platform.claude.com/docs/en/models/opus-5-5/overview was released a week later, it costs $4 per million input tokens, a million output tokens will set you back $20, and there’s no way to disable thinking. System 2 all through and through. Both types of model are just an API call away, so the question becomes when should you use one over the other? This post covers how to use Jev and the next one will compare Jev and Claude head to head. Stay tuned Unlike LLMs that generate the output token by token, Jev generates the entire output in one go. You call Jev with a State JSON object alternatively, a string or array are also allowed and set of typed Questions . Jev then scores every answer to every question in parallel. There are three types of questions with three different specialized kinds of outputs: - Choice - Returns a probability for each option up to 255 and choose the best one. - Score - Places the state on an ordered scale of 2 to 10 levels. - Noul - Returns the probability that a statement is True . Jev is trained using Reinforcement Learning for Calibrated Decisions https://www.sanity.io/glossary/rlcd-reinforcement-learning-for-calibrated-decisions , which, according to Typesafe, guarantees that the output probabilities are properly calibrated: in aggregate, Jev is supposedly right 80% of the time is says that the probability is 0.8. The offical Jev sdk is typesafe sdk https://docs.typesafe.ai/sdk/python and is pretty straightforward to use. Instantiating a TypeSafeClient class with the API key produces a object with the system one method that is the main way to interact with Jev. python import os from typesafe sdk import TypeSafeClient, Noul client = TypeSafeClient api key=os.environ "TYPESAFE API KEY" r = client.system one model="jev-latest", state="Hi, I was charged twice for the same order. Can I get my money back?", questions={"refund": Noul instructions="Is the customer asking for money back?" }, print r.answers "refund" .noul probability of "yes", e.g. 0.98 print r.model, r.usage.input tokens e.g. jev-1.14.0 305 The three required arguments are the model version we wish to use, the state and the set of questions for which we would like an answer. Questions are provided as a dictionary where each value is an instance of a Noul , Choice , or Score object. Each question type is used slightly differently and returns a different set of values. The answer to each question can be found in the answers dictionary of the SystemOneResponse object returned by system one . Noul: one probability Questions of type Noul , cleverly named after Ber- https://en.wikipedia.org/wiki/Bernoulli trial noul- https://en.wikipedia.org/wiki/Bernoulli trial li trials https://en.wikipedia.org/wiki/Bernoulli trial , produce a single probability value that quantifies the likelihood of the answer to the question being True . query = ”Hello- It is urgent I get a refund on this product. The merchant has not been “ “helpful, what can I do now to get this resolved ASAP?” r = client.system one state=query, questions={”refund”: Noul instructions=”Is the customer asking for money back?” }, model=”jev-latest”, pin “jev-1.14” once you tune thresholds r.answers ”refund” .noul 0.98 r.usage.input tokens 305, for $0.0000128 r.model ‘jev-1.14.0’ When the boundary between the two possible outputs, True or False is not clear, we can supply the model with further information using the criteria argument: Noul instructions=”Is the customer asking for money back?”, criteria={”true”: “The customer wants a refund, credit, or reversal”, “false”: “The customer wants information, or a fix to something broken”}, Unlike classic LLMs, there are no other paramters to set: no temperature, system prompt, few-shot examples, etc. Choice: a distribution across a list of possiblities When you have more than just two possible answers to your question, you can use the Choise question type where you can specify arbitrary classes along with the criteria used to decide how to choose among them. Each class is identified by a unique dictionary key both for the sake of establishing the decision criteria as for identifying which probability corresponds to each group. GROUPS = { “cards”: “Anything about a physical or virtual card: ordering, delivery, activation, loss, or a card being declined”, “transfers”: “Sending or receiving money, including transfers that failed or have not arrived”, “fees and refunds”: “Charges, fees, exchange rates, refunds, and reversals”, “account”: “Identity, verification, personal details, app access, and closing the account”, } r = client.system one state=query, questions={ “group”: Choice instructions=”Which team should handle this query?”, criteria=GROUPS } Each call to Jev comes with a confidence value. Confidence is a measure of how split Jev is across all the options it has, based on the probabilities is computed. In the case of a Choice question, the confidence is: so it just measures how far away the highest probability is form an even split across all options. For a Score question, the calculation is slightly more complex https://docs.typesafe.ai/confidence as it needs to take into account how far from the next level a value sits, but the meaning is the same: how sure is Jev that this is the correct answer. In addition to the confidence level, you can also inspect the full set of probabilities so that you can tell not only what the model chose, but also if there were any runner-ups, etc. {"type": "choice", "choice":"cards", "confidence":1.0, "probabilities":{ "cards":1.0, "transfers":0.0, "fees and refunds":0.0, "account":0.0 } } If you squint a bit, you might think that a Choice with two options True and False is just the same as a Noul and you wouldn’t be too far off… however, Typesafe strongly recommends https://docs.typesafe.ai/confidence sticking with Noul for binary choices and Choice for multinomials. Score: a ranking If instead of a choice among different unsorted options you must select a position in a ranking, then you should use a Score . Here the levels are taken to be in the order they appear in the criteria list and identified by the corresponding index. URGENCY = “No deadline. A routine question.”, “The customer wants a fix soon. Nothing is blocked.”, “The customer names a deadline or is blocked right now.”, r = client.system one state=query, questions={ “urgency”: Score instructions=”How urgent is this query?”, criteria=URGENCY } The final score output by the model is identified by a float which corresponds to the weighted average of the various levels. To obtain the corresponding level you can round to the nearest integer. A confidence value and the probabilities assigned to each level is also returned as before. Many questions, one call Jev’s unique structure means that we must reason about it in a different way. Output tokens are free, so ideally you want to get as much output per call as possible. Jev is also intrinsically parallel with the entirety of the processing happening in a single pass. The combination of these two features is a large part of what makes Jev so powerful: You can ask multiple questions about the same state in a single call If you provide a questions dictionary with multiple keys when you call system one Jev will provide the answer to all of them at the same time. One important detail to keep in mind: all questions are processed in parallel, so questions must be independent. There is no way to establish dependencies between one question and another within a single call. There is also only one State object per call, so all questions have to refer to the same State . questions = { “group”: Choice instructions=”Which team should handle this query?”, criteria=GROUPS , “urgency”: Score instructions=”How urgent is this query?”, criteria=URGENCY , “refund”: Noul instructions=”Is the customer asking for money back?” , } r = client.system one state=query, questions=questions This single call costs 501 input tokens, or about $0.000021. Splitting it into three calls, costs 1081 input tokens about $0.000045 and 3x the time. Output really is free. The price of a Jev call is set by the size of your state and your question definitions, not by the answer. Threshold on sureness, not on the label When it comes to automatic decision making based on Jev’s outputs, Typesafe recommends splitting the probabilities into three bands. If it is high enough, you can act automatically, for intermediate levels, ask for confirmation or provide further information and ask again. Otherwise, refer to a human or another system. A 0.05 probability is just as informative as a 0.95 one, so we take the max p, 1-p and threshold on the result python BANDS = 0.9, “auto” , 0.7, “confirm” , 0.0, “human” def band p : sureness = max p, 1 - p return next label for cut, label in BANDS if sureness = cut What Jev cannot do Jev is a fundamentally different class of model and not much is publicly known about its internal architecture. However, the public interface provides important information about its strengths and limitations. Jev can only produce structured JSON output with specific schemas, so tasks such as summarization, explanations, etc are beyond its reach. You must always provide the complete list of options up to 255 options for Choice and 2 to 10 levels for Score , so there’s no way to extract a list of items. It has no support for images or PDF documents and each questions must be independent. No chains of steps, arithmetic or lookups are allowed and it has no support for tool use. It’s also important to keep a few best practices in mind. jev-latest is an alias that can be changed by Typesafe at any time. Pin the model version so a silent version bump doesn’t break the thresholds you tuned. You should also store response.model next to every answer you keep so that you have full reproducibility. If you enjoyed this post, consider subscribing and sharing it with a colleague who needs to make reliable, affordable decisions at scale. Let me know in the comments your thoughts on Jev and how best to use it. The complete, runnable notebook is available in the series GitHub repository. Don’t forget to this post with others who might be interested, and encourage them to so that they have access to the entire backlog of posts and be the first to know when a new a new article is posted.