ENPT

Sage

Last updated:

Beta65–152ms$0.05/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

  • Maturity4.7/10

    Self-serve closed API (POST /decide) with a free plan and paid plans from $14/mo, a public skill/API doc and a pricing page. No open weights, no TypeSafe-compatible /v1/systemone, no OpenRouter or framework integrations found. Small company, no public SLA.

  • Capability8.4/10

    Yes/No, Choice (2–120), Scale, Sort and Tags, plus optional reasoning, web grounding and images (beta). Third-party JevBench v1.6.1: Intelligence 65.6, Calibration 91.6, Capability 78.6, raw p50 0.15 s. Vendor publishes no ECE or calibration method.

  • Adoption5/100

    Launch posts on X and demo repos on GitHub (0–1 stars each). No integrations, third-party apps or verified production cases found (the case-studies page returns 404).

Vendor claims

Read this first

  1. Closed API, beta. No weights, no base model disclosed, no model card. Everything about how Sage works comes from Levanto’s docs; the only independent numbers are Benchmark Heaven’s.
  2. “First decision model” and “#1 on JevBench” are Levanto’s words. Their oldest public benchmark note is dated 23 Jul 2026, but we cannot verify “first”. On the board: Sage was not on JevBench v1.5.7 or the Decision Index (28 Sep). It is ranked on JevBench v1.6.1, where it is #1 on composite and #2 on capability (behind the open-weights Quyet-1.0-Large). On the live board, hosted APIs moved to a separate API leaderboard, where Sage is #1 on both.
  3. Not TypeSafe-compatible. The API is POST /decide with its own request shape, not /v1/systemone. TypeSafe SDKs will not work against it.

What it is

Sage is Levanto Labs’ hosted decision model. You send content plus a typed question; it returns a structured answer with a calibrated probability and no generated text, unless you turn on optional reasoning.

Kind Returns
Yes/No answer (yes/no/null on a near-tie) + probability of yes
Choice chosen (or null if the top two are too close) + per-option probabilities; 2–120 options (20 with an image)
Scale a level on a scale you define
Sort an ordering of items + confidence
Tags which of 1–120 labels apply

Optional extras: reasoning (thinks a question through; output tokens are billed), grounding (web search, $0.10 per search), images (beta), and /decide/batch. Responses report meta.model (levanto-sage-v1.3 in current docs), latency_ms and token usage.

Specs

Attribute Value
Vendor Levanto Labs
Access Hosted API, POST https://sage.levanto.ai/decide, Authorization: Bearer lv_live_…
Weights Closed (base model undisclosed)
Status Beta, self-serve
Earliest public evidence (v0.5 latency note)
Pricing Free plan ($0.20 usage/month); Developer $14/mo … Growth $249/mo; input $0.05–0.038 per 1M tokens by plan; output (reasoning) $10 per 1M; answers are free
Context About 32K tokens per request; 128K on the Growth plan. Oversized requests fail instead of truncating

Third-party board: Benchmark Heaven JevBench (verified 7 Oct 2026)

Board Composite Capability Intelligence Calibration Speed Cost p50 (raw)
v1.6.1 release (all systems) 74.0 (#1) 78.6 (#2) 65.6 91.6 93.6 58.2 0.15 s
Live API leaderboard (board v1.7.x) 74.0 (#1) 78.6 (#1) same scores

For comparison on the same scale, Jev 1.13.0: composite 71.5, capability 77.1. Quyet-1.0-Large (open weights): capability 81.7, composite 71.4. v1.6.x is a full re-measure and not comparable with v1.5.x scores. Sage answered the full 1,500-item set (not equated).

Fit / anti-fit

Fit when you need: a cheap hosted decision call with richer kinds than most peers (Sort, Tags, Scale), abstention on near-ties, and optional reasoning or web grounding.

Anti-fit when you need: open weights or self-hosting, the TypeSafe /v1/systemone contract, a disclosed base model and training method, or an SLA.

Why it is in the catalog

It is a dedicated decision model with a typed API and per-option probabilities, and an independent board ranks it at the top of the hosted field. Approved as a beta closed-API page by Fabricio (7 Oct 2026); same treatment as Liquid d1.

Score working (7 Oct 2026)

  • Maturity = availability 6 × 0.30 + docs 7 × 0.25 + integrations 3 × 0.25 + support 2 × 0.20 = 1.80 + 1.75 + 0.75 + 0.40 = 4.70
  • Capability = types 10 × 0.30 + calibration 7 × 0.25 + latency 8 × 0.25 + accuracy 8 × 0.20 = 3.00 + 1.75 + 2.00 + 1.60 = 8.35 → 8.4
  • Adoption = engagement 8 × 0.35 + ecosystem 4 × 0.35 + production 2 × 0.30 = 2.80 + 1.40 + 0.60 = 4.80 → 5

Accuracy 8 and latency 8 rest on Benchmark Heaven’s measurement (third party), not on Levanto’s numbers. Calibration stays at 7: the board’s calibration axis is strong (91.6), but it is one board and Levanto publishes no method or ECE.

Update 17 Oct 2026: Sage 1.4 and the new API leaders (Benchmark Heaven, board v1.7.46)

Read this first. Sage 1.3.0 is no longer #1 on the API board, and the new Sage 1.4 scores lower than 1.3.0 on the text board. The score of this page stays on the 1.3.0 numbers above (best size and measured longest); 1.4 is documented here, not re-scored.

System (API board, 1,500 items) Composite (rank) Capability Intelligence Calibration p50 (raw) est. $ per 1,000
kettei-fat 77.8 (#1) 86.3 80 93 0.20 s 0.026
Sage 1.3.0 74.0 (#3) 78.6 65.6 91.6 0.15 s 0.0247
Jev 1.13.0 71.5 (#7) 77.1 63.6 90.6 0.24 s 0.0323
Sage 1.4 68.4 (#11) 74.8 65.9 83.7 1.74 s 0.0222

Sage 1.4 has a marginally higher Intelligence (65.9 vs 65.6) but a weaker calibration axis (83.7 vs 91.6) and a p50 of 1.74 s, roughly 11 times 1.3.0’s. The board marks it outside the Jev-class caps (latency 2.8× Jev’s reference), which is what holds its composite down. Cost per 1,000 decisions is slightly lower (est. $0.0222).

Images. On ImageJevBench v0.3.0 (separate board), Sage 1.4 + images (beta) has capability 75.85, second by capability behind kettei-fat (75.95) and #5 on the official composite (47.92: the board’s speed and cost terms pull it down). Same caveat: a different board, never mixed with the text board.

Scores unchanged (maturity 4.7, capability 8.4, adoption 5). Accuracy 8 and latency 8 were set on Sage 1.3.0’s measurement, which still stands (74.0 composite, #3; capability 78.6). The “ranks it at the top” wording in the catalog note above is now historical: as of 17 Oct it is third on composite. Sage 1.4 is not scored separately. A possible re-score is written up in docs/06-quadrante.md for Fabricio to accept or reject; none is applied.

SageCraft: World of Warcraft to level 5 (Oct 2026)

Levanto’s SageCraft is an open-source harness in which Sage picks a dwarf priest’s actions in World of Warcraft from screenshots. It reached level 5 on 6 Oct after 32 sessions, 16 harness fixes and eight recoveries. Only the last session (about 18 minutes, 278 decisions) ran without human input. Levanto reports 4,853 Sage calls costing about $3.32 at public rates, median request 0.8 s, all on levanto-sage-v1.3. A co-founder’s write-up says it is “not a controlled experiment”. A demo, not a benchmark; scores unchanged.