Among September’s many launches, the one developers talked about most came from a startup that was almost unknown two weeks ago. On September 15 TypeSafe AI came out of stealth with Jev, a model that gives up text generation entirely. Within a few days it went viral, Bloomberg covered it as a cheaper and faster alternative to large language models, and an independent catalog of projects built with Jev passed 550 entries in just over a week.
Who is behind it
TypeSafe AI is a San Francisco company founded in 2024 by Diogo Almeida, Erik Gafni and Sasha Sheng. Almeida, the CEO, spent about four years at OpenAI working on RLHF, InstructGPT, ChatGPT and GPT-4. Together with the launch, the company announced a $40 million seed round led by DCVC, which according to Forbes values it at $200 million. The name is a tribute to the economist William Stanley Jevons and his paradox: when a resource becomes more efficient to use, total consumption tends to rise. TypeSafe’s bet is that much cheaper artificial intelligence will end up in many more places inside software.
How it works: state in, typed decisions out
TypeSafe calls Jev the first of a new class of models it names “System One”, borrowing Daniel Kahneman’s distinction between fast, intuitive thinking (System 1) and slow, deliberate thinking (System 2). A request to Jev contains a state, meaning the context to evaluate (a string, a JSON object or an array of text), and a set of typed questions of three kinds:
- Choice: pick one option from a defined set, for example which team should handle a ticket;
- Score: place the state on a scale of described levels, for example the severity of a bug;
- Noul: evaluate a yes/no statement, returning a probability between 0 and 1.
Every answer comes with the probabilities of all options and, for Choice and Score, a confidence value. Because the possible answers are defined in advance, the model cannot return values outside the schema, and TypeSafe presents this as eliminating hallucinations and type errors. The guarantee, however, covers the shape of the answer: a Choice always returns a valid option, which can still be the wrong one.
All questions are evaluated in parallel, in a single pass, instead of generating one token at a time as LLMs do. The model was trained with a technique called Reinforcement Learning for Calibrated Decisions (RLCD), which optimizes probabilities against real outcomes rather than against the preferences of human raters. The goal is calibration: among the answers given with 90% probability, about nine out of ten should turn out to be correct. TypeSafe has not published the architecture, weights or a technical paper; it describes Jev as a transformer-based model trained only on synthetic data, while some observers suspect it is derived from an open-weight LLM.
The claimed numbers, with due caution
Jev costs $0.042 per million input tokens ($42 per billion), and output tokens are free. Claimed response times range from 70 to 500 milliseconds, with most calls around 100. A support ticket with its questions weighs about 300 tokens: one hundred thousand tickets of the same size cost about $1.26.
TypeSafe claims Jev is 40 to 200 times faster and 40 to 400 times cheaper than frontier models on comparable tasks, with peaks of 193.6x and 444.6x. The company itself adds caveats: those figures come from test workflows built by its own team, may contain bias and represent the high end of what can be expected in the real world. The service also runs on the US West Coast, so from Europe network latency should be measured before promising response times to users.
Limitations to know about
Jev does not write replies, summarize documents or explain its reasoning: anything that requires text still needs an LLM. It accepts text input only, it is unreliable with arithmetic, counting and date comparisons, which belong in code, and it reads instructions literally, so ambiguous questions produce worse answers. State and questions share a limit of about 64,000 tokens, and a Choice can have at most 255 options.
Access is also still limited. On September 20 TypeSafe opened sign-ups to everyone, only to pause them two days later because of demand. Jev remains available through Vercel’s AI Gateway at the same price.
Where it fits in agents
The most natural use of Jev is as a “smart if statement”, in the places in software where you need a judgment that hand-written rules cannot express. In agentic systems there are three main cases. The first is routing: deciding whether a request needs a powerful, expensive model, whether a fast model is enough or whether it should go to a person; LangChain has already published a middleware that uses Jev to pick the model based on criteria set by the developer. The second is guardrails: before an agent runs a tool call or a terminal command, Jev can classify it as read-only, reversible or irreversible, and low confidence becomes the signal to ask a human for confirmation. The third is verification: checking whether an LLM-generated answer is consistent with its source, or whether a task is really complete.
Early production reports are encouraging, though self-reported. Metaview, a recruiting platform, says it shipped Jev into all of its agents over a weekend, with candidate searches about ten times faster at the same accuracy.
What it means for anyone orchestrating AI models
Jev is best seen as a component that complements LLMs. In agentic pipelines, a sizable share of calls to a frontier model goes to small decisions: classifying, routing, scoring, verifying. The possible savings depend on how large that share is. If a company spends €8,000 a month on AI calls and 40% goes to decisions of this kind, moving them to a path that costs 5% of the current one saves about €3,000 a month, 38% of the total. If decisions account for 10%, savings cannot exceed 10% even if those calls became free.
For European companies there is also the question of data residency: TypeSafe’s infrastructure is in the United States, and the direct service offers zero data retention only to enterprise customers.
The underlying idea is very interesting for anyone who, like us at ModelHive, works on orchestrating different models: use the right tool for each step, with a generative model where reasoning or writing is needed and a fast, cheap decision model where a choice is enough. We will keep following Jev and the next System One models, which may soon come from the large labs as well.
Leave a Reply