Microsoft-Decision-1: what it is and what it costs

Microsoft-Decision-1 is a small Microsoft model that picks from a list of options you give it and returns a probability for each one. It never writes free text. It's available now in Microsoft Foundry and costs $0.042 per million input tokens, with output free.

Satya Nadella announced it on X on Friday, October 9. His post opens with "Introducing Microsoft-Decision-1, our new model for fast decision-making," and claims it beats both regular LLMs and other decision models on speed and quality. He says Microsoft is already testing it in-house on incident response, quality control and scientific discovery. The post passed 1.2 million views within a day.

What Microsoft-Decision-1 does

You send it a question and a fixed set of answers, and it sends back a calibrated score for each answer. TestingCatalog's writeup lists three question formats: yes or no, multiple choice, and ratings. It can also grade an AI response or an agent's action against a rubric you write. The use cases Microsoft names are mostly sorting work, things like routing requests, classifying tickets, setting priority, checking data, labeling, and deciding whether an agent should take its next step.

The model under the hood isn't new. @testingcatalog reported that Microsoft post-trained Qwen3.5-9B, Alibaba's open-weight model, and plans to move to other bases later, OpenAI's models among them. AI researcher @deliprao noted that it does a single job and handles text only. Images aren't supported.

What it costs

The price is $0.042 per million input tokens, and output tokens are free. For scale, 800 customer emails of about 300 tokens each come to roughly one cent. OpenRouter access is listed as "coming soon," so Foundry is the only way in today.

Microsoft's benchmark claims

These figures come from Microsoft and haven't been checked independently yet:

• Highest accuracy across 36 benchmarks, nearly 150,000 questions that were held out of training
• 4.5 times faster than the runner-up, Quyet-1.0-Large, and 35 times faster than GPT-6 Sol at median latency
• Decisions changed only 1.3% of the time when the same request was reworded or the options were shuffled
• Xbox Research labeled more than 10,000 player feedback items at quality comparable to GPT-6 Sol, 14 times faster and at 200 times lower cost

The consistency number is the one I care about. A sorting tool that gives a different answer when you rephrase the question is hard to trust with real work, and early testers of OpenAI's Decisions API found that its answers changed depending on how a question was worded. If Microsoft's 1.3% holds up in other people's tests, that's the claim worth watching.

The X crowd wasn't fully convinced by the charts. @xeophon joked that Microsoft had "solved chart crimes" and asked why the comparison skipped the models people actually use. Another reply asked Microsoft to run it on JevBench, the benchmark from the company that started this category.

The fourth decision model in a month

TypeSafe launched Jev in mid-September. Perplexity followed on October 2 with its Decisions API at $0.04 per million input tokens, and OpenAI opened its own to every developer on October 6. Microsoft is now the fourth, priced almost exactly like Perplexity.

@scaling01 posted "pretty interesting how fast everyone deployed decision models." Ethan Mollick put it in four words: "The distillation cycle continues." I think that's the right reading. A 9-billion-parameter open model, trained to do one narrow job well, is now matching frontier models on that job at a fraction of the price. The category is filling up fast, and prices are already close to zero.

What it means for a business using AI

Most of the AI work a small business could use is sorting. Is this lead real? Is this message urgent? Which tech should get this job? Does this invoice look wrong? Those are multiple-choice questions, and there are now four vendors selling tools built for them at prices that barely register.

At this point the vendor matters less than the questions. Someone has to write down what each decision is, what the options are, and what a wrong answer costs. If you'd like help finding the sorting jobs in your week that are worth handing off, New Face Design's free process audit starts there.