This New AI Doesn't Chat, It Decides

TypeSafe AI has released Jev, an AI model that makes decisions for software instead of writing text.
The company had been working in secret. It announced Jev on September 15, 2026, in San Francisco, describing it as a new frontier model that brings machine-native intelligence directly into software systems. TypeSafe emerged from stealth with $40M in funding Yahoo Finance.
The founder is Diogo Almeida. He was an OpenAI researcher who helped build ChatGPT and invent reinforcement learning from human feedback (RLHF). Almeida left OpenAI two years ago to start TypeSafe AI TechCrunch.
Jev is not a large language model. It does not write sentences. It gives odds for preset choices, described as calibrated decisions. Users set those choices in advance, so it cannot hallucinate. The model is named after 19th-century economist William Stanley Jevons TechCrunch.
TypeSafe calls Jev a System One model. On its homepage, the company claims Jev is 193.6x faster and 444.6x cheaper TypeSafe AI.
Pricing works the same way. Output tokens are free. Input tokens are counted by the billion, not the million. That flips normal LLM pricing. It treats running the model as measurement, not writing.
Demand was immediate. TypeSafe briefly could not serve users from its API because demand for Jev was so high.
The broader context here is a change in how this AI design is used. For the past several years, production use has meant asking a text generator for an answer and then cleaning up that text to fit into software. Jev skips that step. The developer lists the fixed outcomes, the model returns odds for each one, and the software acts on the numbers rather than words.
In my view, the key claim is not speed or cost, though both matter when decisions happen at large scale. It is calibration. Like a weather forecast that says 80 percent chance of rain, an 0.82 result should be right about 82 percent of the time. Most LLMs used as judges or routers are poorly calibrated out of the box, and teams add thresholds, retries and human review to compensate. If Jev stays calibrated across tasks without extra training for each job, it removes a lot of extra code. If it needs careful setup and monitoring to stay calibrated, it trades one kind of tuning work for another.
Looking at what this means for builders, the model fits control planes, fraud checks, routing, triage and policy enforcement. These are places where delays are measured in milliseconds and costs are measured in fractions of a cent. Free output tokens and per-billion input counting fit that workload. With no text to parse, there are no made-up fields to catch. The risk shifts to badly chosen outcome lists and changes in incoming data.
Worth flagging is how this changes reliability. TypeSafe makes hallucination impossible by design, because there is no open text where false details can appear. That is a useful limit. It does not remove error. A calibrated 70 percent decision is still wrong 30 percent of the time, and software using those decisions needs a clear plan for uncertainty. The benefit is that uncertainty becomes a number that can be tested.
Looking ahead, what this enables, if the early developer response holds, is machine learning that acts like a simple function call. Software already relies on fixed tools for logic. A fast, cheap, calibrated tool for messy judgments could sit next to them for choices too complex for simple rules and too narrow for chat. That is a quieter goal than conversational AI, but it touches far more code.


