Microsoft (Foundry) · Decision model · Last verified 2026-10-15

Microsoft-Decision-1

BetaPartly verified

Small decision model in Microsoft Foundry that returns typed numeric decisions (noul / choice / score) about text or JSON state. No free-form output or rationale.

What it does

  • Endpoint: <foundry-resource-endpoint>/providers/microsoft/v1/systemone. Request "model" is your deployment name.
  • Question types: noul (probability 0–1), choice (one option + probability per option), score (probability-weighted average of ordered level indexes).
  • State may be text or JSON; multiple questions can share one request.
  • Auth: Microsoft Entra ID (recommended) or API key. Deployment types: DataZoneStandard (selected regions) and GlobalStandard. Model version "1".
  • Also available via OpenRouter (provider: Azure) and Vercel AI Gateway.

Pricing

  • $0.042 per 1M input tokens; output $0 (OpenRouter listing, provider Azure). [Verified 2026-10-15 on OpenRouter; Foundry/Azure price page not checked]

Limits

  • 32,768-token context window (OpenRouter). [Verified 2026-10-15 on OpenRouter]
  • Rate limits are per deployment quota (429 on exceed); specific numbers not published on the how-to page. [Verified 2026-10-15 against official page]

Status

  • Released Oct 9, 2026 (OpenRouter). OpenRouter reported 99.99% availability over 3 days on 2026-10-15. [Verified 2026-10-15 on OpenRouter]
  • OpenRouter: "Weights are updated continually while the API shape stays the same." Microsoft docs list model version 1. [Verified 2026-10-15; whether Foundry pins weights per version is unclear]
  • Preview label and any SLA. [Unverified — not confirmed on a Microsoft page]

Safety & fit notes

  • Good fit: gates, routing, ranking, grading, and agent/tool selection inside a fixed allowlist.
  • Microsoft lists known risks: bias, wording/order sensitivity, calibration strongest on familiar tasks, no explanations, may miss harmful content.
  • Microsoft: for consequential decisions about people (credit, employment, housing, healthcare, legal) use it for decision support with meaningful human review, not as the sole decision-maker.
  • Recommended: include a "cannot tell" option, randomize option order, and set thresholds from labeled data.

Compare

Sources