Microsoft Releases Decision-1 for AI Decision Tasks

Microsoft released Microsoft-Decision-1 on Friday, October 9. Rather than generating open-ended text, the model selects from a fixed group of choices and assigns a calibrated probability to each option. That output is designed for software that needs to act on a decision immediately.

The model is available in Microsoft Foundry and is expected to arrive on OpenRouter later. Microsoft lists pricing at $0.042 per million input tokens, with no charge for output tokens.

Microsoft positions Decision-1 as a low-cost control layer for AI agents. These systems often make a series of small decisions between larger steps, such as choosing a model, deciding whether to retry a task, or determining when to send work to a person. Each additional decision can add both time and cost.

How Microsoft Decision-1 Works

Microsoft trained Decision-1 from Qwen3.5-9B, an open-weight model from Alibaba. The company further trained the base model to score available options in one pass.

Decision-1 supports several structured decision formats:

  • Yes-or-no questions
  • Multiple-choice questions
  • Rating tasks
  • Evaluation of AI responses against a rubric
  • Evaluation of agent actions against a rubric

Microsoft also plans to recreate the model using other foundations, including Microsoft’s MAI models and OpenAI models.

According to Microsoft, decision latency matters when choices are made sequentially. Achint Srivastava, a vice president in Microsoft’s Office of the CTO, noted that adding 100 milliseconds to each of 20 consecutive decisions results in two extra seconds for a workflow.

Microsoft Decision-1 Benchmark Results

Microsoft says Decision-1 delivered the highest accuracy in its comparison across 36 benchmarks. The evaluation included nearly 150,000 questions that were excluded from training.

In Microsoft’s measurements, Decision-1 was also the fastest model tested. The company reported that it was:

  • 4.5 times faster than Quyet-1.0-Large, the next-fastest model measured
  • About 35 times faster than OpenAI’s GPT-6 Sol at median latency

Microsoft also tested how consistently the model handled eight variations of the same prompt, including reordered or reworded versions. On average, the model changed its answer 1.3% of the time.

Safety testing included 5,250 requests across 11 benchmarks.

These benchmark results come from Microsoft’s own testing and have not been independently verified.

Internal Microsoft Uses of Decision-1

Microsoft says it is testing Decision-1 across the company for incident response, quality control, and scientific discovery.

Several internal teams have reported results from their own trials:

  • Xbox Research used the model to classify more than 10,000 items of player feedback into themes. The team reported quality comparable to GPT-6 Sol, at more than 14 times the speed and one two-hundredth of the cost.
  • Microsoft Copilot found the model competitive with GPT-5.6 Luna for judging response quality, while reporting a 100-times speed advantage.
  • Microsoft Discovery reported that Decision-1 scoring was 46 times more consistent than an approach based on a large language model.

Microsoft wrote that cost has become a major factor in how people decide to use AI as agentic AI becomes more common.

Decision-1 Availability and Pricing

Microsoft-Decision-1 is available through Microsoft Foundry. Microsoft has also said the model will be available through OpenRouter later.

Its stated pricing is:

 

Usage

 

 

Price

 

 

Input tokens

 

 

$0.042 per million tokens

 

 

Output tokens

 

 

Free

 

The model’s structured output, single-pass option scoring, and probability-based results are aimed at AI workflows where many individual decisions can accumulate delay and expense.

FAQ

What is Microsoft Decision-1?

Microsoft Decision-1 is a model designed for structured decision tasks. It chooses from a fixed set of options and provides a calibrated probability for each choice rather than producing open-ended text.

What tasks can Microsoft Decision-1 handle?

The model can handle yes-or-no, multiple-choice, and rating questions. It can also assess AI responses and agent actions using a rubric.

Where is Microsoft Decision-1 available?

Microsoft-Decision-1 is available in Microsoft Foundry. Microsoft says it will come to OpenRouter later.