TypeSafe AI launches Jev, a non-LLM 'System One' transformer that returns calibrated decisions instead of text

TypeSafe AI, founded two years ago by ex-OpenAI RLHF inventor Diogo Almeida, has released Jev, a transformer-based 'System One Model' API that outputs typed, calibrated probabilities instead of generated text. The hosted early-access product claims major…

TypeSafe AI, a startup founded two years ago by former OpenAI researcher Diogo Almeida, who helped invent RLHF, has released Jev, a transformer-based model the company calls a 'System One Model.' Unlike a large language model, Jev does not generate text: it returns typed decisions with probabilities and confidence values — what the company calls 'calibrated decisions' — making it fast, cheap, and hallucination-free by design since users define outputs in advance. The API exposes three question primitives — Choice (up to 255 options), Score, and Noul (binary probability) — processed in parallel against a shared state, and Almeida says the model was trained exclusively on synthetic data via 'reinforcement learning from calibrated decisions,' which MarkTechPost renders as Reinforcement Learning for Calibrated Decisions (RLCD). Performance claims differ by source: vendor-run benchmarks cited by MarkTechPost claim 193.6x faster and 444.6x cheaper responses than GPT-5.6 Terra — figures the outlet notes are unverified and measured against reference answers averaged from other models — while customers cited by TechCrunch reported Vercel seeing 5-18x faster results than ChatGPT Luna 5.6 for command-safety classification and Bryo AI finding Gemini slightly more accurate but 10-20x more expensive for email classification. Jev is in hosted early access behind a waitlist with no published weights, parameter count, or self-hosting option; its architecture is undisclosed, and outside observers suspect it builds on an open-weight LLM. Demand was reportedly high enough to briefly knock out the company's API.

• TypeSafe AI was founded two years ago by former OpenAI researcher Diogo Almeida, who helped invent RLHF.
• Jev is a transformer-based 'System One Model' that is not an LLM; it outputs calibrated probabilities and confidence values instead of generated text, making it unable to hallucinate by design since users define outputs in advance.
• The API exposes three question primitives — Choice (up to 255 options), Score, and Noul (binary probability) — processed in parallel against a shared state.
• Almeida says Jev is trained exclusively on synthetic data via 'reinforcement learning from calibrated decisions' (RLCD, per MarkTechPost); the architecture is undisclosed with a parallel sampler, and outside observers suspect it builds on…
• Pricing is $42 per billion input tokens ($0.042 per 1M) with output tokens free.
• Performance claims vary by source: vendor-run benchmarks claim 193.6x faster and 444.6x cheaper than GPT-5.6 Terra (unverified, with reference answers averaged from other models), while customer Vercel reported 5-18x faster results than…
• Jev is in hosted early access behind a waitlist, with no published weights, parameter count, or self-hosting option.
• Demand was high enough to briefly knock out TypeSafe AI's API.