TypeSafe came out of stealth on Wednesday with Jev, a model built for software that needs a decision rather than a paragraph.
The company was founded by Diogo Almeida, an OpenAI veteran credited as a ChatGPT co-inventor, and spent two years building the system before launch. Jev takes an unstructured state as input and returns type-safe structured values in a single query, skipping the autoregressive token generation conversational models rely on. Almeida describes it as a frontier-intelligence function call.
Training is the differentiator. Instead of optimising for human preference or verifiable answers, TypeSafe trained Jev to produce calibrated probabilities over execution logic, so confidence attached to a result tracks how often it is right. Outputs are constrained to a predefined schema, which removes parsing layers and guardrail wrappers from the application path, and a parallel sampler scores up to 255 discrete options before choosing one.
TypeSafe’s own measurements put end-to-end latency between 70 and 500 milliseconds from its West Coast servers, against the 3 to 329 seconds it recorded in frontier conversational deployments. On published branching workflows, it reports execution up to 193.6 times faster than consensus baselines from GPT-6 Astra and Fable 5.1. Input processing is priced at $0.042 per million tokens, with output tokens unmetered.
Early developer access opened Wednesday. The targeted uses are real-time feature extraction, verification layers and branching logic where hand-written rules turn brittle.
