Sage2: Levanto’s streaming decision model and Extract

Sage2 on 2 November 2026

Levanto Labs’ launch post says software can understand human language and act on it in real time. The post splits Sage2 into Extract and Decide. Extract takes natural language and returns structured data. Levanto says that path is 166 times cheaper than Anthropic’s Opus 5.5 and beats Cloudflare’s Clairic. Decide is described as the first decision model with streaming, three times faster because it reads context as the text grows. Sign-ups in the next 24 hours were offered 20 million free tokens. The same thread points at the model page, a playground, and an MPP setup guide for agents.

Akshay Pachaar’s same-day write-up frames the streaming path against a decision model that re-bills the full context on every call. In a 40-message chat checked after each message, early tokens would be processed 40 times. Sage2, in that account, keeps a server-side session closer to a database cursor: new messages append, cached prefix computation is reused, and every registered question is re-evaluated against the full history. Each question returns a typed answer with a calibrated probability. On progressive context, Akshay repeats Levanto’s 3x speed claim and places responses in the tens of milliseconds. He also notes that Levanto launched its first decision model before Jev.

Sources:[227] Levanto Labs[231] Akshay Pachaar[229] Levanto Labs

Extract fields and streamed decisions

The product page shows three request shapes. POST /decide takes a yes/no, choice, scale, tags, or sort question over content and returns a structured answer with a probability. The refund demo prints yes at probability 0.999 in 104.1 ms of server time. WS /v1/stream opens a session, registers questions once, and appends text as it arrives. The escalation demo moves from no at 0.009 after a single sign-in complaint, to yes at 0.991 when an EU workspace fails, then yes at 0.999 when checkout fails for every customer, with server times of 110.8 ms, 76.9 ms, and 131.2 ms. POST /v1/extract maps “I can pay $450/mo” to a money field monthly_limit of 450 USD in 182.1 ms.

The page also lists multimodal image analysis, about 150 ms extraction on short documents, and a 3x faster claim for decisions on progressive context, with streaming listed as low as 50 ms per decision. Sessions are capped at 16,384 tokens of text and questions. Extra switches include reasoning (off, auto, or on), web grounding, vision in beta, probability thresholds in application code, and batch requests. Existing Sage keys work with extraction and streaming. Levanto has not printed a full per-token price table beyond the launch free allowance in the sources opened for this page.

Sources:[229] Levanto Labs[227] Levanto Labs

Extract Bench v0.1

Levanto’s launch thread says extraction models were new enough that the team could not find a benchmark, so it published Extract Bench. On that board, Sage2 ranked first in accuracy at 77.8 and in identifying non-answers at 72.2, ahead of Cloudflare Clairic at 75.3 and 59. The open repository at commit 5fb31b8 matches those figures for Sage 2 (rc9 preview, fast) on version 0.1.

Extract Bench scores typed field extraction from real documents. A system receives document text and typed fields (text, number, money, date, duration, or list) and returns a value or null when the document lacks one. Version 0.1 covers 1,844 documents across 12 public datasets of invoices, receipts, forms, contracts, reports, emails, and questions. The README leaderboard also lists NuExtract3 at 76.1 / 57.0, Schematron-3B at 59.9 / 54.3, and GLiNER2.5 at 35.7 / 44.1. A footnote states that Levanto maintains the benchmark and sells Sage 2. Levanto fitted Sage 2’s abstention thresholds on the fit half; the published score uses the eval half only. We have not re-run the bench.

Sources:[228] Levanto Labs[230] Levanto Labs

Normies, Perp Flow, and GrowBlocks

Normies quote-tweeted the launch and wrote that its agents had made over 15 million decisions on Sage in about a month. Normies Arena is described as a fully autonomous agent game played 24/7. The team says it recently switched to Sage2 and was looking for 150–200 ms response time in a real-time game.

Perp Flow wrote that it had used Sage for some time as an auto trader making weekly decisions and to organise a database and find patterns. In the past month alone, Perp Flow reports more than 500,000 decisions while reducing AI usage. Levanto replied that decision models were still not ideal for execution management, and that they can unlock new speeds of quantitative analysis.

The Pond Street Ledger’s launch write-up collects those two posts and adds GrowBlocks City, where one Sage stream per player drives resident reactions, mayor ticket routing, apology checks, and plaza chat screens. The Ledger discloses that its publisher founded GrowBlocks and that GrowBlocks had early access to Sage2. Counters on growblocks.xyz/sage on Monday evening of launch day showed 9,624 decisions that day and 24,478 so far in November, with an average of 406 ms per decision and resident calls in the live feed at 83 to 136 ms. Those figures are self-reported and dated to that snapshot.

Sources:[232] NORMIES[233] Perp Flow[234] The Pond Street Ledger

Next to Jev on this site

jevmodel.ai’s decision-models page already puts TypeSafe’s Jev next to language-model judges and encoder classifiers. Sage2 sits in the same category of typed answers with probabilities, with two product differences that show up in Levanto’s materials: a dedicated Extract path for structured fields, and a streaming Decide session that appends context instead of resending the full history. Akshay’s post explicitly compares that session shape to a decision model that recomputes the whole chat on every check.

This page does not rank Sage2 against Jev. There is no shared public pairing table in the sources opened here. Extract Bench is an extraction leaderboard Levanto maintains. Normies and Perp Flow volumes are those authors’ counts. Official Sage product and docs stay at levanto.ai. Official Jev product stays at typesafe.ai.

Sources:[227] Levanto Labs[231] Akshay Pachaar[230] Levanto Labs[232] NORMIES[233] Perp Flow[229] Levanto Labs

Questions about Sage2