A new kind of AI model from a ChatGPT inventor is thrilling developers

Diogo Almeida, an OpenAI alumnus who helped develop reinforcement learning from human feedback, left the company two years ago and founded TypeSafe AI to build a different kind of model. This week TypeSafe launched Jev, a transformer-based system that deliberately avoids producing natural language and instead returns calibrated probability outputs for predefined tasks. The company positions Jev as a cheap, fast alternative to large language models for software automation. By not emitting free-form text, Jev removes a major source of hallucination and lets users define the precise outputs they expect. TypeSafe says output tokens are free while input tokens are metered on a much coarser scale (by the billion rather than the million), which the company argues reduces cost and latency. Demand for the API surged immediately after release, briefly exceeding the company’s capacity to serve users. Developers testing Jev report significant performance and cost advantages in practical use cases. Vercel, which had previously used OpenAI’s ChatGPT Luna 5.6 to classify commands for safety, replaced Luna with Jev and observed results that were five to 18 times faster and more accurate. In another comparison, Bryo AI’s CTO found Gemini marginally more accurate for classifying business emails but 10 to 20 times more expensive; he highlighted Jev’s real probability scores as especially useful for automating workflows. TypeSafe describes Jev as a 'System One model' focused on intuitive decision-making for the right task rather than general reasoning. Almeida says the model is trained entirely on synthetic data and uses a method he calls 'reinforcement learning from calibrated decisions.' Observers suspect Jev may be built atop an open-weight LLM, but TypeSafe has not disclosed detailed architecture. Potential applications cited by developers include using Jev to monitor or gate LLM agents, route workloads to appropriate models in real time, and provide a low-cost check against misbehavior or hallucination in agentic systems. The model’s name references economist William Stanley Jevons, whose paradox describes how falling costs can spur greater use of a commodity. Almeida envisions Jev’s lower cost and speed leading to more widespread, distributed deployment of intelligent software rather than concentration in large, centralized apps. Outside experts expect competitors to emerge if the pattern of utility and cost savings proves repeatable.