TypeSafe AI’s Jev Brings Faster, Cheaper Automation Without Hallucinations
TypeSafe AI's Jev offers faster, cheaper software automation with calibrated probabilities, free outputs and no hallucinations from undefined responses.
Summary
TypeSafe AI released Jev during the week of September 18, 2026, two years after former OpenAI researcher Diogo Almeida left to found the startup. Almeida helped build ChatGPT and invent reinforcement learning from human feedback, but concluded that four years of optimizing human language had not made LLMs ideal for automation. Jev is a transformer based System One model that returns predefined, calibrated probabilities instead of text, preventing undefined hallucinated outputs. Output tokens are free, inputs are metered by the billion rather than million, and launch demand briefly overwhelmed its API.
Vercel engineer Pranit Sharma said Jev ran a command safety classifier five to 18 times faster and more accurately than OpenAI’s ChatGPT Luna 5.6. Bryo AI CTO Nikhil Mudholkar found Gemini slightly more accurate at classifying business emails, but 10 to 20 times costlier; Jev uniquely returned actionable confidence scores. Earendil CTO Armin Ronacher said users must treat 50 percent confidence as a coin toss and 95 percent as actionable, while Jev could cheaply route workloads, monitor LLM agent traces and prevent jailbreaks. Its undisclosed architecture may use an open-weight LLM, but Almeida says Jev is trained exclusively on synthetic data through reinforcement learning from calibrated decisions, with half of TypeSafe AI operating as a synthetic-data lab. Named for 19th-century economist William Stanley Jevons and his cost-driven consumption paradox, Jev is intended to spread cheap intelligence across software. TypeSafe AI plans new modalities, and Ronacher expects competitors.
Positives
- Vercel measured five to 18 times faster command classification with greater accuracy than OpenAI’s ChatGPT Luna 5.6.
- Jev’s predefined outputs prevent it from generating undefined, hallucinated responses.
- Output tokens are free, while input usage is metered by the billion tokens rather than the million.
- Gemini cost 10 to 20 times more than Jev in Bryo AI’s business email classification test.
- Jev’s calibrated confidence scores can support workflow automation, model routing and jailbreak monitoring.
Risks & concerns
- Gemini was slightly more accurate than Jev in Nikhil Mudholkar’s business email classification test.
- Launch demand briefly exceeded TypeSafe AI’s capacity to serve users through its API.
- Users must decide whether confidence levels, such as 50 percent or 95 percent, justify automated action.
- Jev’s architecture remains undisclosed, while outside observers suspect an open-weight LLM foundation.
- Competitors are expected to emerge now that Jev’s utility is apparent.