Sage2 is the model Levanto Labs is shipping for developers who need application code to act on emails, chats, tickets, and live conversations while they are still arriving. Developers define the fields they need. Extraction returns typed values for text, numbers, money, dates, times, and durations, with the source span and a probability, or null if that detail was never stated. Decision calls answer yes/no, choice, scale, sort, and tag questions and return null when confidence is too low, so the running program can route, escalate, or hand off instead of guessing. Streaming is what this release adds to the decision model Levanto introduced in July 2026. A session holds context on the server, the client appends only new text, and up to 32 questions are scored again as the text grows, within a 16,384-token limit. An earlier answer can reverse when later messages change the evidence. In the launch demo, a car-shopping agent fills vehicle and budget fields as the customer talks, then updates whether to prepare a quote once the customer accepts an offer. Levanto reports that decisions on this progressive context are 3x faster, with streaming as low as 50 ms, and that extraction is 166x cheaper than Opus 5.5. On Extract Bench, a benchmark the company released with the launch, Sage2 scored 77.8 on accuracy and 72.2 on identifying non-answers, compared with 75.3 and 59 for Cloudflare's Clairic. The new extract and stream endpoints work with existing Sage API keys. Sage2 is also available on NEAR AI and, for agents, through MPP. Levanto Labs was founded by Marco De Rossi and Christopher Kocurek.