TL;DR: Microsoft introduced Microsoft-Decision-1 on Oct. 9, 2026, a decision-scoring model in Microsoft Foundry that returns probability scores for fixed options instead of generating text.

Microsoft is pitching a different kind of AI for software that has to act fast. On Oct. 9, 2026, Achint Srivastava, vice president of software engineering in the Office of the CTO at Microsoft, announced Microsoft-Decision-1, a model for fast decision-scoring available in Microsoft Foundry and through OpenRouter, according to Microsoft Command Line. The company described it as purpose-built for routing, classification, prioritization, verification, and workflow control, with structured outputs that software can immediately act on.

The distinction matters because Microsoft framed decision models as a new category separate from large language models. According to Microsoft Command Line, large language models are designed to generate text or reason through complex problems, while decision models are designed to deliver structured outputs at very low cost with high performance. Coverage in The Verge, posted Oct. 9, 2026 at 10:21 PM UTC by Richard Lawler, summarized the same idea: the point is to choose something quickly and efficiently, like figuring out which AI model to use or what is in an image being analyzed.

What Microsoft-Decision-1 actually does

Microsoft-Decision-1 does not write prose or explain its reasoning in open-ended form. When given a fixed set of answer options, it provides a calibrated probability score for each option through a simple structured API call, according to Microsoft Command Line.

The supported formats are broad but bounded. Microsoft said the model supports yes/no, multiple-choice, and rating options, as well as rubric-based grading of AI responses and agent actions. Reporting in MarkTechPost added that the Foundry model card lists yes/no, multiple-choice, rating, classification and rubric questions, grading of AI responses and proposed agent actions, groundedness checks against supplied evidence, and explicit abstention options such as “cannot tell,” with output in JSON.

The practical pitch is decision intelligence inside existing applications, agents, and workflows in a secure, trusted environment. An application supplies the situation and choices, the model scores them, and the application uses confidence to decide when to act, defer, or ask for review.

Built on Qwen3.5-9B, with a rebase planned

To build Microsoft-Decision-1, Microsoft said it post-trained Qwen3.5-9B for fast, single-pass decision scoring and will soon rebase it on other models, including Microsoft AI (MAI) and OpenAI, according to Microsoft Command Line.

That base-model choice drew attention because the Qwen model family is developed by Alibaba Cloud, the cloud computing arm of Chinese tech giant Alibaba, according to The Register. The Register reported that Microsoft said it will soon rebase Decision-1 on its own models and those from OpenAI, while MarkTechPost reported that training used public datasets processed through Microsoft’s Open Data process plus synthetic data.

Microsoft has not disclosed exact parameter counts beyond the Qwen3.5-9B base, and MarkTechPost described the offering as hosted API only through Microsoft Foundry and OpenRouter via Azure, with no open weights. The listed context window is 32,768 tokens for text inputs, according to MarkTechPost.

Accuracy claim across 36 benchmarks

Microsoft said Microsoft-Decision-1 achieved the highest accuracy in a 36-benchmark comparison spanning nearly 150,000 questions across benchmarks kept blind from training, according to Microsoft Command Line. MarkTechPost specified the total as 147,137 questions across nine compared systems.

In Microsoft’s figures, Microsoft-Decision-1 led on average accuracy at 83.5%, ahead of Quyet-1.0-Large at 81.9%, H2O-Lightning-4B at 77.2%, and GPT-6 Luna Decisions at 79.4%, according to MarkTechPost and The Register. The Register also reported that Microsoft ranked the model second behind Quyet-1.0-Large on confidence score, at 92.2%.

Microsoft said it also tested several top public models from the open leaderboard JevBench across 36 additional public and private benchmarks covering routing, ranking, long context, multilingual and out-of-distribution tasks, reasoning, and safety. According to Microsoft Command Line, Microsoft-Decision-1 performed best across those broader sets, which the company presented as evidence of strong generalization.

Speed claim: 4.5x and 35x faster in company tests

Speed is central to Microsoft’s argument because sequential agent steps compound delay. Microsoft gave the example that adding just 100 milliseconds to each of 20 sequential decisions adds two seconds to the overall workflow, according to Microsoft Command Line.

In Microsoft’s benchmarking, the company said Microsoft-Decision-1 was the fastest measured: 4.5 times quicker than Quyet-1.0-Large, the runner-up, and 35 times quicker than GPT-6 Sol. Microsoft Command Line separately stated that Microsoft-Decision-1 P50 latency is about 35 times faster than GPT-6 Sol.

MarkTechPost reported specific latency figures from Microsoft’s chart: p50 latency of 85 ms with p95 at 125 ms, compared with 380 ms median for Quyet-1.0-Large, 210 ms for H2O-Lightning-4B, 300 ms for GPT-6 Luna Decisions, and 3.01 seconds for GPT-6 Sol. The Register reported related comparisons of 2.5 times faster than H2O-Lightning-4B and 2.8 times faster than Jev in Microsoft’s latency test.

Robustness and calibration testing

Microsoft emphasized that equivalent inputs should produce equivalent decisions even when wording or formatting changes. The company said it perturbs the same request in eight ways and measures how often the decision flips, according to Microsoft Command Line.

The reported result was that Microsoft-Decision-1 changes its decision on 1.3% of perturbations on average, with zero flips when option descriptions are paraphrased or when options are reversed or shuffled. The eight-way test covers paraphrased states and instructions, changed option descriptions, reordered choices, changed keys, and formatting noise, which Microsoft said should not materially alter the decision.

Calibration is exposed as part of the product. According to Microsoft Command Line, the probability itself is part of the API, not just a ranking score, so a 90% prediction should be right about nine times out of 10 on representative cases. Applications are expected to use that confidence to choose among acting, deferring, or asking for review.

Safety evaluation on 5,250 requests

Microsoft said it tested Microsoft-Decision-1 on 5,250 requests across 11 benchmarks covering harmful content, jailbreak attempts, and prompt injection, according to Microsoft Command Line. The company reported that the model successfully refused harmful behavior while retaining a high degree of utility.

Microsoft did not publish a single safety score in the announcement excerpt. The framing was that a decision model should recognize harmful requests without needlessly blocking harmless ones, preserving utility for routing and classification work while refusing disallowed actions.

Price and where developers can get it

Microsoft listed input tokens at $0.042 per million tokens with output tokens free, according to The Register and MarkTechPost. MarkTechPost illustrated Microsoft’s modeled estimate that classifying one million texts would cost about $11 with Decision-1 versus about $2,434 with GPT-6 Sol, based on identical briefs and including output charges for the general model.

Availability language differed slightly by source. Microsoft Command Line described the model as available in Microsoft Foundry and through OpenRouter. The Register described it as offered via Microsoft Foundry and, soon, via OpenRouter. The Foundry listing identifies Microsoft as publisher and positions the service for decision and classification workloads.

Early internal pilots at Xbox, Copilot and Discovery

Microsoft described three internal testing efforts. Xbox Research used Microsoft-Decision-1 to process more than 10,000 open-ended pieces of feedback and reviews from surveys, STEAM, and Twitter/X and sort them into a fixed set of themes established by researchers, according to Microsoft Command Line. Microsoft reported quality competitive with GPT-6 Sol while running over 14 times faster and 200 times less expensive.

The Copilot team, which measures quality of chat and agentic responses, found Microsoft-Decision-1 competitive with GPT-5.6 Luna and 100 times faster in its testing, according to Microsoft Command Line. For incident response, Microsoft said on-call engineers use AI to retrieve relevant knowledge across logs, ticketing systems, calls, messages, and other data sources.

MarkTechPost added that in Microsoft Discovery, rubric scoring was 46 times more consistent than LLM-based scoring at three times the speed, while adaptive replanning ran nearly four times faster, as reported by Microsoft.

Why Microsoft launched now

The launch arrived amid a crowded field. The Register traced interest to Jev, announced by TypeSafe AI three weeks earlier as a language model tuned to respond with a limited range of probability-rated responses, and listed OpenAI’s Decisions API in public beta, Cloudflare’s Clef, Strands Decider 2B, Liquid AI’s d1, Perplexity’s Decisions API, Snowflake’s decision model, Surogate Rune, and H2O.ai’s H2O-Lightning-4B built on Qwen3.5-4B, with more than 100 such models vying for attention.

The Verge placed Microsoft’s announcement just days after OpenAI highlighted its new Decisions API, describing the moment as a shift from large language models and agents to decision models. For Microsoft, Srivastava’s message was cost and fit: now that agentic AI is a reality, cost plays a major role in how people use AI, and choosing the right model for the right job matters increasingly, according to The Register.

Sources referenced