Money and business in the Middle East.

AI

Microsoft ships Decision-1.1, twice as fast at same cost

Microsoft made Microsoft-Decision-1.1 generally available, a small model built to pick actions, not generate text.

· Last verified: 20 Nov 2026 (Microsoft Command Line blog, Azure AI Foundry Blog)

Summary

  • Microsoft made Microsoft-Decision-1.1 generally available, a small model built to pick actions, not generate text.
  • The model runs tasks in 123.25 milliseconds, more than twice as fast as its predecessor at near-identical cost.
  • Enterprises can offload routing, triage and escalation calls from costly large language models to a cents-per-million-token engine.

The latest

A decision-making model that chooses among developer-defined options rather than writing text is now generally available, Microsoft said, pricing it at 0.042 dollars per million input tokens globally with output tokens included. The company said Microsoft-Decision-1.1 completes a task in 123.25 milliseconds against 274.25 for the first version, at 139.64 dollars per million tasks versus 142.56.

Details

  • What it does: According to Microsoft, the model was trained from first principles to select the correct outcome from a fixed option set — routing, classification, prioritization, verification, workflow control, tool selection, escalation decisions, incident triage and document classification — instead of invoking a generative model at every step.
  • Calibrated confidence: Microsoft said the model operates inside developer-set constraints and returns calibrated confidence scores, so applications can defer a decision when uncertainty rises. That abstention behavior is the mechanism the company positions as making enterprise automation dependable rather than merely fast.
  • The architecture: The model is built on Gemma 4 26B A4B, a mixture-of-experts system. Multimodality is new in version 1.1: it accepts photographs and screenshots as base64-encoded PNG and JPEG images, not URLs, sized between 280 and 1,568 pixels per dimension. Microsoft did not state the context window.
  • Benchmark scores: Microsoft reported 68 points on the Artificial Analysis Decision Index v2.1 — measuring quality, calibration and latency — with its chart showing 67.78, six points above Decision-1 at 62.07. The same chart places Jev 1.14 at 66.23, Gemini Decision 1 at 65.98 and GPT-6 Luna (Decisions) at 60.25.
  • Accuracy testing: Across 102 benchmarks covering more than 385,000 held-out questions, Microsoft put average accuracy at 78.8% and said the model led seven of nine capability categories. Its table lists 78.83% for the model, 78.12% for Jev 1.14.0, 77.73% for Gemini Decision 1 and 73.4% for GPT-6 Luna, with Jev leading on knowledge and abstention.
  • Cost comparison: On routing Copilot requests, Microsoft said the model is 2.2 times faster and 4.4 times cheaper than a large language model. Rival entries in its chart cost far more per million tasks: 345.41 dollars for Gemini Decision 1 and 338.74 for GPT-6 Luna, against 139.64 for Decision-1.1.
  • Adoption so far: More than 4,000 customers have used the model since last month's public preview, Microsoft said, naming Epic, Equinor, Atos Group and Raiffeisen Bank International among early adopters. Garrett Adams, Epic's senior vice president of research and development, said decision models have enormous potential to make its AI agents more responsive and scalable.
  • Pricing and access: Consumption pricing runs 0.042 dollars per million input tokens globally, 0.046 in the United States data zone and 0.050 in the European Union and Asia-Pacific zones, with no charge for output tokens. The model is available in Microsoft Foundry's catalog and through OpenRouter and Vercel.
  • Deployment setup: Microsoft said the model runs by default in Foundry Evaluations for performance, quality and safety, and is called with Entra ID authentication without an API key, requiring Python 3.10 or later and the Cognitive Services User role.
  • Internal use: Microsoft said it uses the model to route requests between Copilot and Cowork, that Cowork's Auto mode relies on it to tune reasoning effort per task, and that on-call engineers use it for context retrieval during incidents. The company frames it as a companion to generative models and its own model router.

Background

Microsoft announced the release on its Command Line blog on 17 November 2026 in a post by Ashish Srivastava, vice president of software engineering in the Office of the CTO, and on the Azure AI Foundry Blog on 18 November 2026 in a post by Saumil Shrivastava.

Between the lines

The pitch targets cost structure more than capability. By pairing a 4.4-times cost advantage on Copilot routing with calibrated abstention, Microsoft argues the expensive part of agent architectures is calling a generative model for decisions that need no generation. Epic's stated interest in making agents more scalable suggests early demand sits in high-volume, low-creativity steps.

What's next

Watch whether the 4,000 preview customers convert to production deployments, whether Epic moves from evaluation to live clinical use, and whether independent Artificial Analysis Decision Index results match Microsoft's reported 68-point score.

Source: Microsoft Command Line blog, Azure AI Foundry Blog