On Friday, October 9, 2026, Microsoft introduced Microsoft-Decision-1, an artificial intelligence model that is not designed to chat or write, but to decide. Instead of generating text, it reads the content it is given and returns a calibrated probability for each of the fixed options the developer defines. The announcement was written by Achint Srivastava, vice president of software engineering in Microsoft’s Office of the CTO, and Satya Nadella himself shared it on social media.
The company describes decision models as an emerging category. Unlike large language models, which are designed to generate text or reason through complex problems, these systems produce structured outputs that software can act on immediately. Microsoft argues that once you understand this ability to classify and decide at very low cost and high performance, many uses open up: routing requests, prioritizing tasks, verifying answers or controlling an agent’s workflow.
Technically, Decision-1 is a post-trained version of Qwen3.5-9B, Alibaba’s open model, tuned for single-pass decision scoring. Microsoft says it will soon rebase it on other base models, including those from its own Microsoft AI (MAI) division and from OpenAI. It supports yes-or-no, multiple-choice and rating questions, as well as rubric-based grading of AI responses and agent actions, all through a structured API call.
The main argument is speed. Microsoft explains that every decision adds delay, especially when one step depends on the previous one: adding just 100 milliseconds to each of 20 sequential decisions adds two seconds to the overall workflow. In its tests, Decision-1 was the fastest model measured: 2.5 times faster than H2O-Lightning-4B v1.1, the runner-up, and about 35 times faster than GPT-6 Sol, OpenAI’s flagship model, in median latency.
On quality, the company says it achieved the highest accuracy in a comparison across 36 benchmarks totaling nearly 150,000 questions, all kept out of training. The benchmarks cover routing, ranking, long context, multilingual and out-of-distribution tasks, reasoning and safety. Microsoft says it took several of the top public models on the open JevBench leaderboard and evaluated them on those 36 additional public and private benchmarks, to check that quality was not limited to a single type of task.
The company pays particular attention to robustness. Equivalent requests should produce equivalent decisions, even if the instructions are paraphrased, the order of the options changes or formatting noise appears. Microsoft perturbed each request in eight different ways and measured how often the decision flipped: Decision-1 did so in 1.3% of cases on average, and in none when option descriptions were paraphrased or options were reversed or shuffled.
Calibration is another key piece, because the probability is part of the product. Applications use the model’s confidence to decide when to act, when to defer or when to ask for human review, so a 90% prediction should be right about nine times out of ten. On safety, Microsoft tested the model on 5,250 requests from 11 benchmarks covering harmful content, jailbreak attempts and prompt injection, and says it refused harmful behavior without needlessly blocking legitimate requests.
The company shares examples of internal use. Xbox Research used Decision-1 to sort more than 10,000 pieces of feedback and reviews from surveys, Steam and X into a fixed set of themes, and found it competitive in quality with GPT-6 Sol while running more than 14 times faster and costing 200 times less. The Copilot team, which measures the quality of the assistant’s responses, found it comparable to GPT-5.6 Luna and a hundred times faster. And the Microsoft Discovery science platform uses it to replan experiments based on rubrics.
Pricing strengthens the pitch: $0.042 per million input tokens, while output tokens are free, since the model returns only a short answer. According to its OpenRouter listing, it supports a 32,768-token context. The model does not work with the usual chat interface but with a dedicated decisions endpoint, and the listing itself states that it is not intended for open-ended generation, conversation, translation or summarization. Microsoft updates its weights continually while keeping the same interface.
The launch confirms that the industry is moving toward this kind of specialized tool. On October 6, OpenAI opened its Decisions API, based on GPT-6 Luna, in public beta, and there are alternatives such as TypeSafe AI’s Jev. The reason is practical: as companies chain together agents that make hundreds of small decisions, using a large, slow model for every classification is expensive and inefficient. A small, calibrated model can handle those steps and leave complex reasoning to the larger systems.
Microsoft suggests a long list of applications: controlling agents by evaluating whether their next step should continue, stop, retry or be handed to a person; choosing the best model for each request based on quality, cost and latency; labeling data; judging other models’ responses; classifying incidents by type and urgency; filtering content; reviewing code; or selecting the next action in a graphical interface or a robot.
There are some caveats. All the performance figures come from Microsoft itself, which also added comparisons with Jev to its post after publication, and independent evaluations are only beginning. The fact that the model is built on a Chinese open system may also weigh on some organizations, although Microsoft has already said it will switch the base to its own and OpenAI’s models. And continual weight updates mean users need to watch for possible changes in behavior.
Even so, Decision-1 illustrates an underlying trend: with agents acting in the real world, the cost and latency of each step matter as much as the intelligence of the main model. “It’s increasingly important to choose the right model for the right job,” Microsoft says. If decision models deliver on their promise, they could become the nervous system of automated workflows, handling the thousand small choices an agent has to make every minute.
