TypeSafe’s Diogo Almeida launches System One Models with JEV for structured decisions

Diogo Almeida, TypeSafe’s founder, is betting that the real gap in AI isn’t better small talk. It’s automation. After two years in stealth, his company has introduced Jev, the first System One Model in a new line aimed at software that needs fast, structured decisions instead of generated prose.

Almeida frames the launch as a response to a problem he’s been thinking about since his OpenAI days, when he worked on the methods that made language models follow instructions and talk with people. That research fed into ChatGPT. But, he says, the chat breakthrough still left something big missing. TypeSafe’s answer is a new stack built around a new model architecture, a parallel sampler, and a training method it calls Reinforcement Learning for Calibrated Decisions, or RLCD.

Jev is the first public model from that effort and is available now in early access. TypeSafe says it reaches similar intelligence to existing large language models on System One tasks while being two orders of magnitude faster and more efficient. The tradeoff is deliberate: Jev gives up string generation, because it is designed around type-safe structured outputs and calibrated probabilities rather than free-form text.

That design choice runs through the whole pitch. TypeSafe says Jev never makes type errors, always reports confidence, and can slot into ordinary software for jobs like classify, route, score, extract, or branch. The company also says its structured outputs avoid hallucinated tool calls, which matters a lot more once AI sits inside code paths with latency guarantees and dependency chains.

The launch material is full of comparisons to traditional LLMs, including claims about cost, speed, and workflow performance. TypeSafe says Jev’s input pricing is $0.042 per million tokens, output tokens are free, and end-to-end responses can land between 70ms and 500ms. It also says early-access demos and workflow tests show gains that reach the company’s headline figures, though some of the supporting evaluations rely on its own setup and reference models. For now, the big idea is simple: stop asking models to be chatty when what software really needs is dependable decisions.