Prince Mario-Max Schaumburg-Lippe: Microsoft Decision-1 Scores Choices in 85 Milliseconds

Most AI models love to talk. Microsoft Decision-1 does not. Announced on October 9, 2026, it skips the chat entirely, reads a question with a fixed list of answer options, scores every option in a single forward pass, and hands back a calibrated probability for each one. That is the whole product. No paragraphs. No small talk. Just a decision, with numbers a piece of software can act on right away.

Microsoft built the model by post-training Alibaba’s open-weight Qwen3.5-9B. It runs as a hosted API only, through Microsoft Foundry and through OpenRouter, with a 32,768-token context window and text-only input. Microsoft says it plans to rebase the model over time on other foundations, including its own MAI models and OpenAI systems. For now, the Qwen base is the engine, and Microsoft is open about that.

A different job than chatting

Think about how much machine work is not writing at all. A support ticket needs routing to the right team. A photo needs a yes or no on whether it fits a policy. A claim needs sorting into urgent, routine, or check again. An AI agent, mid-task, needs someone to verify its last step before it takes the next one. Those are choices from a short, known list, repeated thousands or millions of times a day.

Using a big chat model for that work is a bit like hiring a novelist to stamp forms. It can do it. It will also write you three friendly paragraphs around the stamp, slowly, at full price. Decision-1 is the stamp. Microsoft lists the target jobs plainly: routing, classification, prioritization, verification, and workflow control, plus grading AI responses and agent actions against a rubric. The supported question shapes are yes or no, multiple choice, and ordered ratings, where the returned score is a probability-weighted average across the levels.

That focus connects to a wider shift in how teams build with smaller, specialized models. Our recent coverage of five small Qwen models teaming up to match a frontier giant made the same point from another angle: a fleet of right-sized models, each doing one job well, can beat one giant doing everything at once.

The numbers Microsoft published

Microsoft supports the launch with its own benchmark run, and the figures are worth reading with that source in mind. Across 36 benchmarks covering about 150,000 held-out questions (147,137 in the detailed write-ups), Decision-1 averaged 83.5 percent accuracy, the highest average in Microsoft’s comparison. The next name in that table, Quyet-1.0-Large, came in at 81.9 percent. On calibration, Decision-1 scored 92.2, a close second to Quyet-1.0-Large at 93.1.

Speed is where the pitch gets loud. Microsoft reports a median latency of 85 milliseconds, with the 95th percentile at 125 milliseconds. It says the model runs about 35 times faster than GPT-6 Sol on its own test, 4.5 times faster than Quyet-1.0-Large, and 2.5 times faster than H2O-Lightning-4B v1.1. Stability held up too: in Microsoft’s testing, small rewordings of the input flipped the chosen option only rarely, with decisions staying the same 98.7 percent of the time.

Price helps the case. Input costs $0.042 per million tokens, and output tokens are free, which makes sense for a model whose output is a handful of probabilities rather than a page of prose. Microsoft frames that as more than 20 times cheaper than GPT-6 Sol for text classification work. Every one of those numbers is vendor-reported, from Microsoft’s own testing, so treat them as a strong opening claim rather than a settled league table. Independent checks will land in time. The direction, though, is hard to argue with: decisions are cheaper when the model stops writing essays.

Why agents need a referee

The most interesting use is the quiet one. As AI agents take longer chains of steps, someone has to check the work between steps. Is this draft good enough to send? Does this tool call match the request? Should this case go to a human now, or can the system carry on? A fast scoring model fits that referee job neatly, because the check itself cannot cost more time than the task it is checking.

That is also where governed, enterprise-grade agent work is heading. Our report on DXC putting governed AI agents to work for banks and insurers showed big organizations asking for exactly this kind of control layer: clear rules, auditable calls, and a human-readable record of why the machine did what it did. A probability score is a record. You can log it, threshold it, and review the borderline cases on a Friday afternoon with a coffee.

What a builder actually sends

The request format is refreshingly plain. You send a state (text or JSON describing the situation), then one or more questions. Each question has a type, instructions, and criteria. A yes-or-no question can carry short descriptions of what true and false mean in your context. A choice question lists each option with its own description. A score question lists ordered levels, and the answer comes back as a weighted value across them. The response returns an answers object, keyed by question name, with the chosen option and a probability per option.

That design has a practical charm. Product teams already think this way. Support desks have routing trees. Moderation teams have policy rubrics. Operations teams have priority ladders. Decision-1 lets those teams hand their existing rubrics to a model instead of rewriting their process around a chatbot and hoping the prose parses cleanly on the other side.

Where it fits, and where it does not

Decision-1 will not write your report, brainstorm your campaign, or explain its reasoning. It returns no text and no explanation at all. If your task needs a reasoned argument a person must read, you still want a language model, or a person. If your task needs ten thousand tidy calls a day, each one logged and priced in fractions of a cent, this is the shape of tool you have been waiting for.

There is a nice honesty in that division of labor. The last two years of AI were dominated by models that tried to do everything in one long answer. The next phase looks more like a well-run office: specialists at each desk, a fast clerk stamping the forms, and the big thinkers called in only when the question actually needs thinking. Voice tools are finding their own desks too, as our story on Block’s Dessa learning to talk from voices alone showed this week.

For teams building today, the takeaway is simple. List the decisions your product makes on repeat. Write each one as a fixed set of options with plain criteria. Then test a scoring model against your current setup on your own data, and watch two numbers: accuracy on the borderline cases, and cost per thousand calls. If Microsoft’s published figures hold up even halfway in the wild, a lot of quiet, repetitive machine judgment is about to get much faster, and much cheaper to run well.

Also published on: