Sage2

Sage2 is a runtime decision and extraction model launched by Levanto Labs on November 2, 2026. You send natural-language text (chat, tickets, document snippets) plus questions or field schemas; the model returns calibrated yes/no or multi-choice decisions and structured extractions with source spans. A WebSocket streaming session keeps context server-side and updates answers as new text arrives—Levanto claims streaming decisions are about 3× faster than re-scoring the full history, with ~150 ms extraction on short documents and streaming decisions as low as ~50 ms.

It is not the same product line as Sage-A1.5 (the InternScience research model already in this database). Sage2 is meant as an application-level primitive for support, sales, trading and game loops. Product pages, docs and a playground are live, and Levanto advertises about 20M free tokens for new sign-ups.

Core capabilities

• Decide: define yes/no or multi-choice questions in plain language ("Does this ask for a refund?", "Escalate to on-call now?") and get answers with probabilities for routing tools or ticket severity.
• Extract: declare typed fields (money, dates, durations, text); Sage2 returns values with source spans and probabilities.
• Stream: open a session, append only new text, reuse prefix compute server-side; emit every update or only when answers change.
• Multimodal: Levanto says Sage can also analyze images.
• Extract Bench: Levanto's own extraction benchmark claims 77.8 accuracy and 72.2 on identifying non-answers, ahead of Cloudflare Clairic (75.3 / 59); the benchmark is open-sourced on GitHub.

Integration

• Docs and the API contract are written so coding agents can wire it in; existing Sage keys work for extraction and streaming.
• Also available via NEAR AI confidential inference and agent-facing MPP (with Stripe / Tempo).
• Case mentions include Leap, Capacitr and PerpFlow.

Latency, cost multiples and leaderboard numbers are vendor claims; independent replication is still limited.

Who it is for

• Product teams that need sub-second routing and field extraction inside ongoing support, sales, trading or game chats.
• Agent builders who want a cheap, specialized decide/extract primitive instead of calling a frontier LLM for every yes/no.
• Compliance-minded extraction workflows that need citable source spans.

Getting started

1. Visit levanto.ai/model and playground.levanto.ai, sign up and claim the trial credit.
2. Read docs.levanto.ai and try POST /v1/extract plus streaming WS /v1/stream on real conversation snippets.
3. Wire probabilities into business rules or tool-selection logic; fall back to a general model or a human on low confidence.
4. Optionally extend Extract Bench with your own domain eval set.