Jev returns typed decisions instead of generated text
TypeSafe's Jev returns a yes/no probability and a 0–1 confidence score for bounded tasks, completing a demo task in 0.114 seconds, positioned as a cheap decision layer in front of larger LLMs.
𝗔 𝗖𝗵𝗮𝘁𝗚𝗣𝗧 𝗖𝗼 𝗜𝗻𝘃𝗲𝗻𝘁𝗼𝗿 𝗝𝘂𝘀𝘁 𝗟𝗮𝘂𝗻𝗰𝗵𝗲𝗱 𝗝𝗲𝘃 𝗔𝗻 𝗔𝗜 𝗧𝗵𝗮𝘁 𝗠𝗮𝗸𝗲𝘀 𝗗𝗲𝗰𝗶𝘀𝗶𝗼𝗻𝘀 𝗜𝗻𝘀𝘁𝗲𝗮𝗱 𝗢𝗳 𝗪𝗿𝗶𝘁𝗶𝗻𝗴 𝗔𝗻𝘀𝘄𝗲𝗿𝘀 Jev comes from TypeSafe, founded by Diogo Almeida, who worked on training methods behind ChatGPT. What makes Jev different is that it does not generate text word by word. Give it a bounded task and it returns a typed decision, a yes or no probability, and a confidence score between 0 and 1. In one launch demo, Jev completed a task in 0.114 seconds. A frontier LLM took 8.6 seconds on the same task. That makes it interesting for jobs where you do not need an AI to write anything. Which team should receive this email? Is this customer a good fit? Is this comment spam? Which model should handle this request? Jev can make the decision first. A larger LLM such as Claude only needs to step in when actual generation or deeper reasoning is required. The economics are also different. TypeSafe says processing one million messages can cost around $16.50, depending on the input. Its API is currently in early access, but Jev can already be accessed through Vercel, OpenRouter and Cloudflare. The important limitation is confidence. A clean output does not mean a correct decision. Low confidence cases should still be routed to a human or a stronger model. That may be where Jev becomes most useful: not as a replacement for LLMs, but as a cheaper decision layer in front of them. Which AI workload would you move to something like this first?