A model that writes no text: TypeSafe releases Jev
On 15 September 2026 TypeSafe AI, founded by the former OpenAI researcher Diogo Almeida, released Jev, a model that generates no strings. Instead of text it returns typed structured values with calibrated probabilities. The set of possible answers is defined in advance, so the model cannot step outside the schema.
Why it matters
For four years the industry pushed the quality of human language at the output. Here is the opposite move, offered commercially for the first time: give up language and take speed, price and the impossibility of leaving a predefined schema in exchange. Whether this is a separate class of models or a niche beside LLMs is open: the company does not disclose the architecture, and outside observers suspect an open-weight model underneath.
The announcement came on 15 September 2026 and TechCrunch covered it on the 18th. Almeida founded the company two years ago after leaving OpenAI, where he helped build ChatGPT and helped develop reinforcement learning from human feedback. The company calls Jev's training method Reinforcement Learning for Calibrated Decisions and says the model is trained exclusively on synthetic data. The figures the company states: input tokens at 0.042 dollars per million and output tokens free; end-to-end response times of 70 to 500 milliseconds against 3 to 329 seconds for frontier LLMs, which it puts at 40 to 200 times faster; and on its own workflow evaluations, 193.6 times faster and 444.6 times cheaper. Those evaluations are its own: the reference answer is the average of GPT-6 Astra and Fable 5.1, and the workflows were written by its own staff, as the post's nuance notes say. Independent testimony exists too. Pranit Sharma, an engineer at Vercel, replaced GPT-5.6 Luna with Jev in a classifier that reviews commands for safety and reported results five to eighteen times faster and more accurate. Bryo AI's chief technical officer Nikhil Mudholkar compared it with Gemini: Gemini was slightly more accurate but ten to twenty times more expensive, and what he valued most was the probability on the output. Armin Ronacher of Earendil names the other side: the model delegates the hallucination problem to its user, who now decides what to do with an answer that carries 50 per cent confidence. Demand was such that the company briefly lost the ability to serve its API.