Microsoft Decision-1 is a new AI model that doesn’t write anything. You hand it a fixed set of options, and it hands back a probability for each one, so software can pick a path fast. Microsoft announced it on October 9, 2026, and it’s live now in Microsoft Foundry.
If you build apps or agents, think of it as a cheap, quick referee that sits between “the AI had an idea” and “the app did something.”
- Announced: October 9, 2026, with a post from Satya Nadella.
- Where: Microsoft Foundry now. OpenRouter is listed as “coming soon.”
- Price: from $0.042 per million input tokens, with free output tokens.
- Size: about 9 billion parameters, post-trained from the open Qwen3.5-9B model.
Those details come from TestingCatalog’s report, which also notes that most performance numbers are Microsoft’s own.
What Microsoft Decision-1 actually does
A normal chatbot writes an answer and hopes you like it. Decision-1 skips the writing. You send a question plus a list of choices, such as yes/no, multiple choice, or a rating scale, and it returns a calibrated probability for each choice.
Your code can then act on the score. High confidence, go ahead. Low confidence, retry, escalate, or ask a person.
Microsoft says it also handles rubric-based grading of AI answers and agent actions. That makes it a natural fit for routing support tickets, labeling feedback, screening for unsafe content, or checking an agent’s next step before it runs.
Price: $0.042 per million input tokens, output free
The headline number is low. At $0.042 per million input tokens with free output, you can score a lot of decisions for pennies. For context, we covered OpenAI’s version of this idea in our OpenAI Decisions API guide, which launched at $0.10 per million tokens.
One catch: the report doesn’t say whether free output is permanent or a launch promo. Don’t build a cost model on it until Microsoft says so.
Speed and accuracy claims, and what to check
| Microsoft’s claim | What it means | What to verify |
|---|---|---|
| Top accuracy across 36 benchmarks (about 150,000 held-out questions) | It beat the other models Microsoft compared | Which benchmarks, and how rivals were set up |
| 4.5x faster than the runner-up and 35x faster than GPT-6 Sol at median latency | Decisions come back quickly enough for live apps | Latency on your own prompts and region |
| Decisions changed only 1.3% of the time across eight rewordings | Answers stay steady if you reorder or rephrase options | Whether your messy real inputs behave the same |
| Xbox Research labeled 10,000+ feedback items at 200x lower cost than GPT-6 Sol | Big savings on bulk labeling | Quality on your own labels, not Xbox’s |
Every figure in that table is Microsoft-reported. None were independently verified in the coverage we found, so run your own test before you trust them.
Where it fits, and where it doesn’t
Decision-1 is built for control and classification. Microsoft lists routing, prioritizing, verification, data labeling, AI judging, search relevance, and safety screening.
It won’t draft an email or explain a concept. If you need open-ended text, you still need a regular chat model. The smart pattern is a big model for thinking and a small decision model for gatekeeping.
This also puts Microsoft in a growing lane. Cloudflare moved first with open models, which we broke down in our Cloudflare Clef explainer. Now all three big platforms are selling “score, don’t write” models.
There’s a practical reason for the rush. Using a frontier chat model to answer a simple “is this spam?” question is like hiring a lawyer to read your mail. It works, but it’s slow and costs far more than it should. A small scoring model does that one job in a blink.
The risk runs the other way. A model that returns a clean probability can look more certain than it deserves. Always check whether a “90%” score really is right about nine times in ten on your data.
How to try Microsoft Decision-1 today
You need a Foundry account. Here’s a sensible first test, kept deliberately small.
- Pick one narrow job, like tagging support tickets as urgent or not.
- Send 200 to 500 past examples you’ve already labeled by hand.
- Compare its picks against your labels and check how it handles low-confidence cases.
If it matches your humans, route the easy cases to it and keep people on the borderline ones.
Frequently Asked Questions
What is Microsoft Decision-1?
It’s a 9B-parameter model from Microsoft that scores fixed options with probabilities instead of generating text. It runs through a structured API in Foundry.
How much does Decision-1 cost?
It starts at $0.042 per million input tokens, and output tokens are free. Microsoft hasn’t said if free output is temporary.
Is Microsoft Decision-1 open source?
It’s built on the open-weight Qwen3.5-9B, but we haven’t confirmed that Decision-1 itself is open. One report points to a Hugging Face page under a different name, so treat that as unclear.
Can it replace ChatGPT or Copilot?
No. It only chooses between options you give it. It can’t chat or write.
The bottom line: test it on one boring job first
Decision-1 is most interesting for teams burning money on a big model just to say “yes” or “no.” If that’s you, a one-week pilot on a single routing task is cheap and will tell you quickly if Microsoft’s numbers hold up. If you’re just a curious user, there’s nothing to switch on yet; this is plumbing for developers.
Leave a Reply