Startup TypeSafe AI releases Jev, a model that classifies inputs without generating text

TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, has released Jev, a model that returns classifications and probabilities rather than generating text. The system is designed for developers to define questions and allowed answers, with response times between 70 and 500 milliseconds. The company notes that while Jev is fast and cheap, it does not prevent errors and only guarantees adherence to predefined options.

Jev's architecture departs from conventional language models by skipping sequential text generation entirely, computing multiple classification outputs in parallel. This design yields response times between 70 and 500 milliseconds, with additional questions adding negligible latency. The model's pricing—$0.042 per million input tokens with no output charges—positions it as a low-cost alternative for high-volume classification tasks embedded within existing software systems.

The company's performance claims rest on four self-designed workflows benchmarked against other AI models' responses rather than independently verified ground truth, and the evaluations omit GPT-6 Astra. While TypeSafe emphasizes Jev's inability to hallucinate, that guarantee extends only to output structure—the model can still select a factually incorrect option from the permitted set, leaving error prevention to the surrounding software logic.