Ex-OpenAI Researcher's Jev Model Skips Text for Structured Logic

Diogo Almeida, one of the researchers credited with helping build ChatGPT, has launched a model that deliberately refuses to write prose. His startup, TypeSafe, has released Jev, a system built not to converse but to make fast, type-safe programmatic decisions inside software pipelines.

Almeida left OpenAI two years ago to start TypeSafe. Jev is the company's first public product, and it targets a narrower problem than the chatbots that made his former employer famous: the countless backend calls where software needs a classification, a score, or a structured value rather than a paragraph of text.

Unlike large language models, Jev does not generate tokens one after another. It uses a parallel sampling architecture instead of autoregressive generation, and it was trained with a technique TypeSafe calls Reinforcement Learning for Calibrated Decisions, which optimizes the model to output well-calibrated probabilities on defined execution logic rather than free-form answers.

Because developers specify the possible outputs in advance, TypeSafe argues Jev cannot hallucinate in the way conversational models can. Internal testing put end-to-end latency between 70 and 500 milliseconds, and separate workflow benchmarks reported Jev running up to 193.6 times faster than conversational frontier models on equivalent tasks.

Early Adopters Report Big Speed Gains

Vercel swapped Jev in for OpenAI's ChatGPT Luna 5.6 on a classifier task and says it saw results 5 to 18 times faster, with higher accuracy on that specific job. Bryo AI ran its own comparison, pitting Jev against Google's Gemini on classifying business emails.

Bryo AI found Jev came in 10 to 20 times cheaper than Gemini for the task, even as Gemini edged it out slightly on accuracy.

That tradeoff, cheaper and faster but not always the most precise, captures the pitch behind Jev: it is not trying to out-reason general-purpose LLMs, just to outrun them on narrow, well-defined decisions at a fraction of the cost.

Pricing reflects that positioning. Jev charges $0.042 per million input tokens and does not meter output tokens at all, a structure that would look unusual for a conversational model but makes sense for a system whose outputs are short, fixed-format decisions rather than lengthy generated text.

For engineering teams running high-volume classification, routing, or scoring logic, Jev offers an alternative to bolting a general chatbot onto a task it was never optimized for. Whether that niche is big enough to sustain a standalone model company is the open question TypeSafe now has to answer with paying customers.