Jev Puts AI Judgment Directly Inside Software
TypeSafe AI has released Jev, a transformer-based model built to automate programmatic decisions rather than generate text.
The company exited stealth to launch it. 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, an OpenAI researcher who helped build ChatGPT and invent reinforcement learning from human feedback (RLHF), a training method that uses human ratings. Almeida left OpenAI two years ago to start TypeSafe AI TechCrunch.
Jev is not a large language model. It does not output text. It produces probabilities described as calibrated decisions, where the numbers are meant to match real-world odds. Users define those outputs in advance, so it cannot hallucinate. The model is named after 19th-century economist William Stanley Jevons TechCrunch.
TypeSafe classifies Jev as a System One model, its term for fast, automatic judgment. On its homepage, the company claims Jev is 193.6x faster and 444.6x cheaper TypeSafe AI.
Pricing follows the same logic. Output tokens are free. Input tokens are metered by the billion, not the million. That inverts LLM pricing. It treats inference, running the model, as measurement, not generation.
Demand was immediate. TypeSafe briefly lost the ability to serve users from its API because demand for Jev was so high.
The broader context here is a shift in where the transformer, the core architecture behind modern AI, is applied. For the past several years, production use has meant prompting a text generator and then parsing, validating and constraining its output back into software. Jev collapses that stack. The developer declares a fixed output schema, a preset list of allowed outcomes, the model returns a distribution over those outcomes, and downstream code branches on probabilities rather than strings.
In my view, the interesting claim is not speed or cost, though both matter for high-volume decisioning. It is calibration. A classifier that returns 0.82 should be right about 82 percent of the time at that confidence level. 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 holds calibration across domains without per-task fine tuning, or added training for each task, it removes substantial glue code. If it requires careful specification and monitoring to stay calibrated, it trades prompt engineering for distribution engineering.
Looking at what this means for builders, the model fits control planes, fraud checks, routing, triage and policy enforcement. These are places where latency budgets are measured in milliseconds and unit economics are measured in fractions of a cent. Free output tokens and per-billion input metering map directly to that workload. No text to parse also means no hallucinated fields to catch. The failure mode moves elsewhere, toward mis-specified outcome sets and drift in input distributions.
Worth flagging is how this reframes reliability. TypeSafe's formulation makes hallucination impossible by construction, because there is no open vocabulary to hallucinate into. That is a useful constraint. It does not remove error. A calibrated 70 percent decision is still wrong 30 percent of the time, and software that consumes those decisions needs explicit handling for uncertainty. The benefit is that uncertainty becomes numeric and testable.
Looking ahead, what this enables, if the early developer response sustains, is machine learning that behaves like a function call. Software systems already depend on deterministic libraries for logic. A fast, cheap, calibrated probabilistic primitive could sit alongside them for judgments that are too messy for rules and too narrow for chat. That is a quieter ambition than conversational AI, but it touches far more code paths.
