Most AI products released this year want to talk to you. Jev refuses to. TypeSafe AI's first model, which the company calls a System One model, takes a situation and a list of statements and returns whether each statement is true, with a confidence percentage attached. That is the whole interface.
It is built by Diogo Almeida, a former OpenAI engineer who co-wrote the core training techniques behind ChatGPT, and the performance claims are aggressive. TypeSafe's own figures put Jev at up to 194 times faster and 445 times cheaper than frontier models such as GPT-6 Astra, which the company translates into intelligence-per-dollar it describes as off the charts. Tom's Hardware walked through the design and the numbers behind it.
Why dropping the conversation makes it fast
Two things account for the speed, according to the company. The model is trained with a method it calls Reinforcement Learning for Calibrated Decisions, and it produces structured output rather than prose. Because no question depends on the text generated before it, multiple statements in one request can be evaluated in parallel instead of streamed out token by token.
That is the actual architectural difference. A chat model generates sequentially because each word conditions the next. Jev has nothing to condition on, so the work parallelises.
The practical shape is a plain JSON in, plain JSON out API. Your code hands over a state, meaning the situation and its data, plus the statements to assess. There is no global knowledge base, no retained memory, no system prompt engineering to keep behaviour stable. Parsing inputs, handling dates and reading from your database all stay in your own code, where they were already working.
The worked example in the documentation is a refund flow. Show the model a customer's recent card transactions, some account details and their last message, then ask whether the customer is requesting a refund. You get yes or no with a confidence rating. Above a threshold you pick, say 85%, your code moves on and asks which resolution they want, and Jev returns a probability distribution across refund, store credit and unclear. The logic stays in your application.
The counterintuitive part is the context limit
Jev does not want the full history leading up to the question. Supplying extra information actually reduces accuracy, and the context window is capped at 64,000 tokens, which is small by current standards.
That reads like a weakness until you consider what it implies. Every other model this year has competed on holding more context. A model that gets worse with more context is making a different claim: that the developer should decide what is relevant, and the model's job is narrow judgement on a curated state.
It also changes the failure modes. Because the output always carries a confidence figure, Jev does not hallucinate in the way a chatbot does, inventing facts that were never supplied. It can still misclassify, respond to literal wording instead of meaning, and fall to adversarial input. Responsibility for acting on a given confidence level sits with the developer, not the model, which is a design choice as much as a disclaimer.
The workflows this suits are the ones where confidence thresholds are already how the code thinks. Intent routing, citation checking, and conditional chains of the form: if we are reasonably sure the customer wants a refund, and prefers store credit, and their purchase history clusters in September, offer the discount.
It is not trying to replace the chat models
TypeSafe is direct that open-ended work belongs to an LLM, and expects the two to run together. Its example is a monitoring system where Jev assesses from the supplied state whether something has genuinely broken, and only then hands off to a language model to read logs, find a cause and write the report.
That division is the part worth paying attention to. A large share of production AI spending currently goes on using a conversational model as a classifier, wrapping it in prompt scaffolding to force deterministic-looking output, then paying for the tokens it generates while thinking out loud. If a purpose-built evaluator does that step at a fraction of the cost, the expensive model gets reserved for the work that needs prose.
The claims are unverified, and 194 times faster is the kind of figure that tends to shrink under independent testing. What is easier to assess is the shape of the product. The friction in embedding AI in software has been the amorphousness, coaxing a text generator into behaving like a function. An assessment-and-confidence endpoint is a function. Developers have spent two years building that abstraction themselves on top of chat APIs, which is a decent signal that someone would eventually sell it directly.
The timing also fits a run of stories about models behaving in ways their operators did not intend, from agents moving faster than humans can review them to Claude now leading a quarter of Anthropic's own research. A model that cannot take an open-ended action, and reports how sure it is about the one thing you asked, is a narrower target for that class of problem.