Since TypeSafe AI launched Jev in late September, decision models — specialized AI systems that return structured answers with probabilities instead of text — have drawn interest from multiple major players. AWS, Upstage, Perplexity and Cloudflare now offer hosted decision models, while independent developers have published open weights on Hugging Face. OpenAI has not yet signed on to the emerging standard.
The decision model API reduces every decision to three question types: choice (select from a list), score (place on an ordered rubric), and noul (yes/no, returning a probability). Developers can mix all types in one request. The model never writes explanations or code, returning only structured answers with a confidence value. TypeSafe's documentation notes that confidence measures how concentrated the probability distribution is and advises teams to set thresholds based on their own data.
The pattern echoes the earlier standardization of LLM chat completions, where OpenAI's API became the de facto interface. The company later launched Open Responses, an open specification based on its Responses API.
Practical use cases for decision models have emerged quickly. On September 25, OpenRouter launched Jev Router, using the decision model to choose the most appropriate LLM and reasoning effort for each incoming request, reserving expensive models for tasks that need them.