Microsoft Decision-1 model picks choices without text

Designed for agent routing and task triage, the system builds on Qwen3.5-9B and charges strictly for input tokens.

Most autonomous agent workflows spend unnecessary compute on small mechanical choices. When software needs to pick the next API tool or check whether a script failed, standard architectures prompt a generative model to write out paragraphs of reasoning before landing on an answer. That setup slows down execution loops and piles up operational costs.

On October 9, 2026, Microsoft launched Microsoft-Decision-1 to handle those choices without conversational text. Announced by chief executive Satya Nadella and Microsoft developer teams, the model evaluates structured alternatives directly from input data. It is currently available in public preview through Microsoft Foundry and on OpenRouter.

What it is: single-pass decisions

Microsoft-Decision-1 is built on a post-trained version of the open-weight Qwen3.5-9B architecture. Rather than generating sentences token by token, the model evaluates inputs in a single forward pass. Developers supply input state as text or JSON alongside predefined options: binary yes or no prompts, multiple-choice lists, numeric scales, or rubric criteria.

The API returns calibrated probability scores across those options without explanatory text. Because the model outputs no sentences, developers avoid waiting for token generation to complete. Microsoft stated it plans to adapt this single-pass post-training approach to future model releases from Microsoft AI (MAI) and OpenAI.

How Microsoft Decision-1 operates in practice

Pricing reflects the non-generative setup. Microsoft set the fee at $0.042 per 1 million input tokens and set output token charges to zero.

In performance benchmarks published by Microsoft, Decision-1 recorded a median (P50) latency up to 35 times faster than reasoning models like GPT-6 Sol, and 4.5 times faster than the open-weight model Quyet-1.0-Large. Across 36 blind evaluations spanning roughly 150,000 test cases on routing, policy compliance, and long context, Microsoft Research reported a 1.3% decision shift rate under eight synthetic text perturbations, dropping to 0% when choices were reordered or rephrased. These figures reflect internal vendor benchmarks and have not yet been corroborated by independent third-party evaluations under real production traffic.

What internal deployments show

Microsoft reported results from three internal teams that tested the model prior to the public announcement:

• Xbox Research: classified over 10,000 player reviews gathered from surveys, Steam, and X. The team reported 14 times faster classification and a 200-fold cost reduction compared to GPT-6 Sol.
• Copilot team: used the model as an automated judge to review answer quality, reaching speeds 100 times faster than GPT5.6 Luna.
• Microsoft Discovery: applied Decision-1 to rubric-based replanning in scientific experiments, yielding a fourfold acceleration in planning loops and 46 times higher consistency than general-purpose language models.

Where it fits and where it fails

Decision-1 is intended for narrow decision points where speed and repeatable outputs matter more than conversational flexibility. Practical targets include choosing tool calls in agent loops, triaging incoming tickets, checking compliance rules, and scoring model outputs in automated testing pipelines.

The model is not designed for open-ended writing, summarizing documents, or explaining the reasoning behind its choices. Systems that require conversational responses or creative generation still require conventional generative models.

What remains unverified

As independent coverage from InfoWorld noted, separating decision scoring from text generation tackles cumulative latency in multi-step agents. However, several operational questions remain open. Performance across noisy real-world data may diverge from Microsoft’s synthetic benchmarks, and the company has not provided deployment dates for applying the technique to future OpenAI or MAI checkpoints.

Sources

• Primary: Microsoft CommandLine announcement
• Primary: Satya Nadella announcement on X
• Primary: Microsoft Tech Community (Azure AI Foundry Blog)
• Documentation: Microsoft Learn technical guide
• Independent analysis: InfoWorld report
• Distribution: OpenRouter catalog entry