TypeSafe's Jev Model Hits Major Adoption by Returning Structured Answers Instead of Text

TypeSafe AI, a San Francisco research lab founded by former OpenAI researcher Diogo Almeida, has unveiled Jev, a fundamentally different kind of AI model that sidesteps text generation altogether. Rather than producing prose, Jev answers structured questions by returning typed probabilistic decisions that applications can process and act on directly.

How Jev Works Differently#

Instead of prompting an LLM and parsing its text output, developers submit a state (text or structured data) alongside a set of typed questions. Jev processes all questions in parallel and responds with three primitive answer types—Choice, Score, and Noul—each carrying a probability distribution and a confidence score. This lets calling code apply threshold-based logic: act above a confidence floor, escalate below it.

The model operates at striking speed and cost efficiency: input runs $0.042 per million tokens, with free output; context window extends to 32,000 tokens; and end-to-end latency ranges from 70ms to 500ms. The training approach, which TypeSafe calls Reinforcement Learning for Calibrated Decisions, prioritizes reliability over breadth.

Swift Ecosystem Uptake#

Adoption accelerated rapidly. Vercel integrated Jev into AI Gateway within 24 hours, where it captured nearly 13% of paid teams—double the share of the GPT-5.6 family on launch day. Netlify followed suit, while LangChain released a TypeSafeClassifier integration bundling model routing and safety screening. Five independent Elixir clients appeared within days.

Real-World Performance Gains#

Early testers reported dramatic improvements in speed and expense. Vercel engineer Pranit Sharma measured a safety classifier running 5–18 times faster than the LLM it replaced. Bryo AI's CTO found Jev roughly comparable to Gemini for email classification accuracy but orders of magnitude cheaper, and noted that Jev alone returned calibrated probability values. One analysis of 12,759 launch tweets by tallied median speedups at 7x (against a 193.6x headline claim) and median cost savings of 30x.