Summary
TypeSafe AI has released Jev in early access, a model designed to return typed probabilistic decisions rather than generated text. The company says its parallel architecture delivers 70–500 millisecond responses for supported automation tasks.
TypeSafe AI has released Jev in early access, a model designed to make fast, structured decisions that software can use directly. The company describes it as the first example of a new category called “System One Models”, which accepts application state and returns typed decisions with probabilities instead of generating a string of text.
The announcement, published on September 15, 2026, comes from TypeSafe founder Diogo Almeida, who says he previously helped build instruction-following methods at OpenAI that contributed to the research behind ChatGPT. TypeSafe says Jev is intended for automation tasks such as classification, routing, scoring, extraction and branching inside ordinary software.
A model built for decisions rather than prose
Most large language models generate text sequentially, producing one token at a time. Software can use that text, but normally needs to parse and validate it before acting on the result. TypeSafe’s approach defines the possible output structure in advance. Jev then returns a value that conforms to that schema, along with probabilities and confidence scores.
The model uses a parallel sampler that produces its outputs in a single query rather than generating a sequence of tokens. TypeSafe says this design is paired with a training method called Reinforcement Learning for Calibrated Decisions, or RLCD. The stated goal is not simply to produce an answer, but to provide probabilities that correspond more consistently to the model’s accuracy on a decision task.
The distinction is important for automation. A type-safe output cannot contain an invalid field or an incorrectly formed tool call, according to TypeSafe’s technical explanation. That property concerns the structure of the result; it does not by itself make every classification or probability correct. Decision quality is evaluated separately through the company’s workflow tests.
Jev can process structured program state that includes text, but it gives up general-purpose string generation. This makes it less suited to open-ended writing and more suited to software components where the surrounding code defines the available choices and controls what happens next.
Reported speed, cost and workflow results
TypeSafe reports end-to-end response times of 70 to 500 milliseconds for its service. It compares this with a range of 3 to 329 seconds for the frontier models in its comparison, describing Jev as 40 to 200 times faster for similarly capable, “System One-shaped” queries.
The company lists Jev’s input price as $0.042 per million tokens, or $42 per billion tokens, with output currently free. TypeSafe contrasts this with input prices of $0.20 to $10 per million tokens for the language models in its comparison and output that is approximately five times more expensive than input.
Its more ambitious figures come from workflow evaluations rather than a general-purpose benchmark. Each model receives the same workflow, represented as code, and the results are compared with the average predictions of two large external models identified by the company as Astra and Fable 5.1. TypeSafe reports that Jev reaches the top of its measured speed-and-cost trade-off and says the evaluations produced headline figures of 193.6 times faster and 444.6 times cheaper.
Those results are specific to the company’s workflow design. TypeSafe says the workflows were created by members of its model-capabilities team, which could introduce bias. It also uses a wrapper to make the language models return structured decisions and probabilities, and says this can be slower and more expensive than asking those models for decisions without probabilities.
Where TypeSafe expects Jev to be used
The company positions structured decisions as a way to insert AI into existing application logic. A developer could use a model output as a probabilistic “if” condition, route cases to different processes, score records in a large dataset or apply several small decisions before reaching a final branch.
TypeSafe’s demonstrations include a Doom-playing bot that reacts to structured game state and a Wikiracing system that chooses links while navigating between Wikipedia pages. The company says Jev supports choices with cardinality up to 255; for larger choices, it uses a two-stage process that scores options independently before making an explicit selection.
Jev is currently in early access, with TypeSafe bringing developers off its waitlist. The company says it is seeking feedback on which decisions developers want to automate and where the model performs well or falls short. It also says the long-term sustainability of its current pricing has not yet been established, although it expects prices to decrease rather than increase.