
Decision and extraction models for your app's runtime
Sage2 is a specialized API from Levanto Labs that returns typed decisions (yes/no, choice, scale, tags, sort) and extracts structured data (text, money, dates, durations) from text, images, and streaming sessions — positioned as a fast, cheap alternative to calling a general-purpose LLM at runtime. Every response includes a calibrated probability so applications can act on confident answers and route uncertain ones to review.
Text plus a natural-language question in, a typed answer with probability out. Supports yes/no, choice, scale, tags, and sort question types for routing, classification, and prioritization.
Open a session, define questions once, and append text as it arrives. Sage retains the context server-side and reuses cached prefix computation, so clients send only new text; decisions update in tens of milliseconds (as low as 50 ms per decision).
Turns natural language into typed values — text, amounts, dates, durations — with source spans, character positions, and per-value probabilities. Extracts one value or every matching mention; short-document extraction runs ~150 ms server-side.
Every answer ships with a probability (e.g., P(yes) 0.999), so apps can set thresholds for which answers trigger actions and which go to human review.
Reasoning can be set to off, auto, or on for harder questions, and web search can be enabled when a decision needs current information, with sources returned.
Send an image with optional text and ask yes/no, choice, scale, or tags questions about it.
Batch many questions or records in one request to cut round trips; a copy-paste prompt lets coding agents (e.g., Claude Code) implement integrations using docs at docs.levanto.ai.
Use POST /decide for one-shot decisions, WS /v1/stream for growing context, or POST /v1/extract for structured fields; all three work with the same API key.
You specify the question kind (yes/no, choice, scale, tags, sort) and criteria, or field types like money and duration; Sage returns structured, typed results your code can act on.
Responses include probabilities and latency in meta.usage; use thresholds to decide what runs automatically versus what goes to review, with your application handling tool execution, validation, and fallbacks.
Usage-based plans where the monthly fee includes the same amount of usage credit; one balance covers decisions, extraction, and streaming. Rates: input $0.038–0.05 per 1M tokens by plan, output $1 per 1M tokens, web search $0.10 per search. No seats and no overage invoices on any plan. The free tier includes $1 of usage per month; higher tiers add larger credits, lower input rates, 128K context (G
$0
Includes $1 of usage a month (~50K short questions, ~2.2K invoices extracted); input $0.05 / 1M tokens
$14/mo
Includes $14 of usage a month (~700K short questions, ~31K invoices extracted); input $0.05 / 1M tokens
$49/mo
Includes $49 of usage a month (~2.7M short questions, ~110K invoices extracted); input $0.046 / 1M tokens
$99/mo
Includes $99 of usage a month (~5.9M short questions, ~220K invoices extracted); input $0.042 / 1M tokens; priority support
$249/mo
Includes $249 of usage a month (~16M short questions, ~550K invoices extracted); input $0.038 / 1M tokens; 128K context; priority support
It's a focused API that makes runtime decisions (yes/no, choice, scale, tags, sort) and extracts typed structured fields (text, money, dates, durations) from text, images, and streaming sessions, returning each answer with a calibrated probability.
Per the vendor's published comparisons, Sage returns decisions in a median 330 ms versus ~1,110 ms for GPT-5.6 Terra and ~3,190 ms for Claude Opus 5, at roughly $0.000017 per decision versus $0.00093–$0.0042 for those models. It's built for structured decisions and extraction rather than open-ended generation.
Plans run from Free ($0, includes $1 monthly usage) to Growth ($249/mo, includes $249 usage). Metering is $0.038–0.05 per 1M input tokens by plan, $1 per 1M output tokens, and $0.10 per web search. Example costs: ~$0.00002 per short question, ~$0.00042 per decision on an 8,000-token document, ~$0.00045 per invoice extraction, ~$0.0001 per five-tag moderation call.
Yes. The Free plan includes $1 of usage per month, roughly 50K short questions or 2.2K invoices extracted, with no seats and no overage invoices.
You open a WS /v1/stream session, send your questions once, then append text as it arrives. Sage keeps the conversation or document server-side and reuses cached prefix computation, so your client only sends new text; it can emit every update or only changed answers, and sessions hold up to 16,384 tokens.
Yes, in beta. You can send an image with optional text and ask yes/no, choice, scale, or tags questions about it.
Yes. You can enable web search per call when a decision needs current information; Sage returns the sources it used, and searches are billed at $0.10 each.
No open-source offering is mentioned; it's provided as a hosted API with documentation at docs.levanto.ai.