Microsoft launches Decision-1, a cheap AI model that picks answers instead of chatting

Microsoft has a new kind of AI model, and it does not chat. Microsoft-Decision-1, announced on October 9, is built to pick the right answer from a fixed list of options and attach a confidence score to each one. The company is pitching it as a cheap, fast building block for AI agents, and the launch puts Microsoft in a crowded new race to own what some engineers now call "decision models."

The announcement came in a post on Microsoft's Command Line blog by Achint Srivastava, vice president of software engineering in the Office of the CTO. It drew quick coverage from The Register, The Decoder, Neowin and others, plus a busy thread on Hacker News. Here is what the model does, what Microsoft claims, and why it matters if you build with AI or simply use tools that run on it.

What Microsoft-Decision-1 actually does

Most people think of AI as a chatbot that writes paragraphs. Microsoft-Decision-1 works differently. You give it a question and a set of allowed answers, and it returns a probability for each option in a single pass. It is not meant to write open ended text at all.

That sounds narrow, but a huge share of what software does with AI today is exactly this kind of choice. Should this support ticket go to billing or to tech support? Is this comment spam? Which tool should an agent call next? Does this answer meet the quality bar, yes or no? Microsoft lists uses such as model routing, data labeling, intent analysis, incident routing, content filtering, safety screening, code scanning and choosing the next action in a user interface.

According to Microsoft, the model is calibrated, meaning a 90% score should be right about nine times out of ten in representative cases. That matters because a confidence number is only useful if it means something. A developer can, for example, let the model act alone when it is very sure and send the borderline cases to a person.

Built on an open model from Alibaba

The first version is not built from scratch. Microsoft says it starts from Qwen3.5-9B, an open weight model from Alibaba Cloud, and was post trained for single pass decision scoring. The Register pointed out that Microsoft is leaning on a Chinese lab's model here, and noted that Microsoft plans to rebase Decision-1 on its own MAI models and on OpenAI models soon without saying why.

The speed and price claims

Cost is the headline. Microsoft prices the model at $0.042 per million input tokens, and output tokens are free. Since the model only returns scores for a short list of options, output is tiny anyway, so input is nearly the whole bill.

Microsoft also makes big performance claims, all self reported:

• Highest accuracy in a comparison across 36 benchmarks covering nearly 150,000 questions. The Register put the figure at 83.5% accuracy.
• About 2.5 times faster than H2O-Lightning-4B, the runner up, and roughly 35 times faster than OpenAI's GPT-6 Sol at median latency.
• About 2.8 times faster than Jev, a rival decision model from TypeSafe AI, according to The Register's reading of Microsoft's figures.
• A decision flip rate of 1.3% on average when inputs were slightly altered across eight types of changes, and no flips when answer options were reordered or reworded.

Microsoft also shared internal examples. Xbox Research used the model to sort more than 10,000 pieces of player feedback into fixed themes, with quality the company described as competitive with GPT-6 Sol at more than 14 times the speed and 200 times lower cost. The Copilot team used it to check the quality of chat and agent responses.

Treat the numbers with care

None of these results have been independently tested yet. Microsoft names its comparison models in an appendix, but it has not published the full test data. The Register presented the figures as Microsoft's own and pointed readers to a Cisco blog post on decision model errors, along with research on how those errors can be magnified when they feed into automated systems. Even at 83.5% accuracy, roughly one answer in six would be wrong, which can do real damage if nobody checks the output.

Why the Microsoft-Decision-1 launch matters

The bigger story is the category. The Register counted a long list of companies that have shipped decision models or decision APIs in recent weeks, including OpenAI's Decisions API in public beta, Cloudflare's open weight Clef models, Liquid AI's d1, Perplexity's Decisions API, Snowflake's model and the small open source Strands Decider 2B. It cited a Hugging Face index with more than 100 such models.

The reason is simple. AI agents make many small choices for every task they finish, and paying a large language model to make each of those choices is slow and expensive. "Now that agentic AI is a reality, we've seen that cost plays a major role in how people decide to use AI," Srivastava wrote, according to The Register. A tiny model that answers in milliseconds and costs fractions of a cent changes the math.

If you are new to how agents work, our guide to AI agents and where they still fail explains why these small decisions add up. The price pressure also fits a wider trend we saw when Anthropic cut prices with Claude Haiku 5.5: the cheapest tier of AI is getting much cheaper, fast.

There is a safety angle too. Satya Nadella recently argued that AI systems need deterministic controls and an emergency brake. A model that can only choose from a fixed menu is easier to fence in than one that can say anything, which is part of the appeal for businesses that worry about runaway agents.

Who is affected

Developers and companies building AI products feel this first. Anyone who now pays a big model to classify, route or grade text has a new, much cheaper option to test. The model is available through Microsoft Foundry and OpenRouter, according to Microsoft, with documentation on Microsoft Learn.

Everyday users will mostly notice it indirectly. Decision models sit behind support chats, spam filters, content moderation and the hidden steps inside AI assistants. If they work as promised, those systems get faster and cheaper to run. If they make mistakes with high confidence, people may get routed to the wrong place or have content flagged without a clear reason.

OpenAI is also part of the picture. Microsoft compares its model to GPT-6 Sol, OpenAI's cheaper model, and plans to rebase on OpenAI models later, even as both companies sell competing decision tools.

What builders should do now

If you run AI workflows, a few practical steps make sense:

1. Find the steps in your pipeline that are really multiple choice questions, such as routing, tagging or pass or fail checks.
2. Test a decision model on your own data before switching. Vendor benchmarks rarely match real traffic.
3. Use the confidence score. Let the model act alone only above a threshold you have measured, and send the rest to a person or a larger model.
4. Log decisions so you can audit errors later, especially for anything touching money, safety or account access.

What to watch next

The main questions are independent testing, licensing and the promised rebase. The Register said Microsoft did not state a license for Decision-1, which matters for companies that want to run it themselves. Watch for outside benchmarks, for Microsoft moving the model onto its own MAI base, and for how quickly rivals respond on price. You can follow more model news on our AI page.

Frequently asked questions

What is Microsoft-Decision-1?

It is a small AI model from Microsoft that picks from a fixed set of options and gives each a calibrated probability. It is designed for routing, classification and checks, not for writing text.

How much does Microsoft-Decision-1 cost?

Microsoft lists it at $0.042 per million input tokens, with output tokens free.

Is it better than GPT-6 Sol?

Microsoft says it is about 35 times faster than GPT-6 Sol on decision tasks and competitive in quality on its internal tests. These are self reported results that have not been independently verified.

Where can developers use it?

Microsoft says it is available in Microsoft Foundry and on OpenRouter, with documentation on Microsoft Learn.