What matters in AI.

Subscribe

Microsoft launches Decision-1.1, a model that picks steps for agents

Microsoft tells that Decision-1.1 got 78.8% on average in 102 tests and is more than two times as fast as version 1.

Microsoft made Microsoft-Decision-1.1 available to all customers in Microsoft Foundry, its cloud platform for AI. A decision model does not write text. It picks one result from a set that you give, for example which team gets a customer query. Microsoft tells that the new version is more correct and has more speed than the first, and that it can read images.

What is new

The model uses Gemma 4 26B A4, with more training for decisions and a custom engine. It gives a confidence score. A team can then set when an agent acts alone and when a person does a check. It is also in OpenRouter and Vercel, which send queries to models. Foundry can use it to judge quality and safety.

The numbers

Microsoft did tests with 102 benchmarks (public tests) and more than 385,000 questions. The average was 78.8%. The model was first in 7 of 9 categories.

What is not known

These tests are by Microsoft. The source gives no external test, no speed number and no price. It tells only that the price is by use, with output tokens at no charge.

Insight: Decision models are becoming a gate before large models

In October 2026, more than 1 in 4 queries on the Vercel router went to decision models. Microsoft made Decision-1 available that month on Qwen3.5-9B, and version 1.1 uses Gemma 4 26B A4. This reads as a race for the cheap first step of each agent task. The model that picks that step also decides which large model gets the work. Test one such step first.

Sources

  1. Introducing Microsoft-Decision-1.1: Build faster, more efficient AI applications with decision modeltechcommunity.microsoft.com
AI MATTER · NEWS · AI MATTER · NEWS ·18 NOV2026

Posted

Tags

Learn the terms in this story

More in Models

All Models news