On October 9, Microsoft released Microsoft-Decision-1 in Foundry: a model that never writes a single sentence. Ask it a question with a fixed set of options, and it hands back one probability per option. That’s it. No prose, no explanation, and it does so in about 85 milliseconds. Welcome to a new family: decision models. Spoiler: they’re not here to replace your chatbot, and understanding them is not rocket science.
The signal operator in the control tower
Picture a railway station. The LLM you chat with every day is the brilliant consultant in the lobby: ask anything, get a well-argued page back. Useful, but slow and expensive, and you don’t need a dissertation to decide which track train 4812 should take.
That job belongs to the signal operator in the control tower. They don’t write reports. They look at the situation, and pull one lever among three. Fast, repeatable, thousands of times a day. A decision model is exactly that: a small, specialized model (Decision-1 is a post-trained Qwen3.5-9B, about 9 billion parameters) built to pick an option, not to talk.
The contract is strict. You send a situation (up to 32,768 tokens of text), a question, and a closed list of options: yes/no, multiple choice, a rating scale, or a grading rubric. You get back a probability for each option, as plain JSON numbers. Output tokens: zero. Which explains the pricing: $0.042 per million input tokens, output free.
A probability, not a vibe
The interesting word in Microsoft’s pitch is calibrated. A calibrated 90% should be right about nine times out of ten; a weather forecaster who announces 90% rain and stays dry half the time isn’t calibrated, just confident. That property is what lets software act on the score directly: auto-approve above 0.95, send anything between 0.60 and 0.95 to a human, reject below. The model can even return “cannot tell” if you give it that option, a built-in way to abstain instead of guessing.
A concrete example
A support ticket arrives. One call, three questions:
Situation: Ticket #4812 — "I've been charged twice this month, this is unacceptable!"
Q1 (yes/no) Is it urgent? → yes: 0.91 · no: 0.09
Q2 (choice) Which department? → billing: 0.96 · technical: 0.03 · sales: 0.01
Q3 (scale) Frustration level? → calm: 0.02 · frustrated: 0.31 · very angry: 0.67
Three levers pulled in one pass, each with a confidence attached. Your code routes the ticket to billing, flags it urgent, and bumps its priority because the customer is fuming. No LLM essay was harmed in the process.
That shape fits a lot of the plumbing around AI systems: routing requests to the right model or path, classifying and labeling data, acting as guardrails that score an agent’s proposed tool call before it runs (continue, stop, or hand off to a human), and grading AI answers against a rubric, the job usually called LLM-as-judge.
The numbers (with the asterisks)
Microsoft’s own evaluation covers 36 benchmarks and 147,137 questions:
| Model | Avg. accuracy | Median latency |
|---|---|---|
| Microsoft-Decision-1 | 83.5% | 85 ms |
| Jev 1.13 | 82.3% | 240 ms |
| Quyet-1.0-Large | 81.9% | 380 ms |
| GPT-6 Luna Decisions | 79.4% | 300 ms |
| H2O-Lightning-4B | 77.2% | 210 ms |
The cost argument is the loudest: Microsoft’s chart puts classifying a million texts at about $11 with Decision-1, against about $2,434 with GPT-6 Sol. Internally, the Xbox Research team reports sorting player feedback 14 times faster and 200 times cheaper than with GPT-6 Sol.
Now the asterisks, because honesty is house policy. All these figures are Microsoft-reported, not independently reproduced; measured through OpenRouter, median latency looks closer to 190–270 ms than 85. The original comparison table omitted Jev, the most-used decision model (it was added later in an editor’s note), and the price chart compares against OpenAI’s flagship rather than its cheaper siblings. And on calibration, the metric that matters most for threshold-based automation, Decision-1 actually ranks third (92.2), behind Jev (93.7) and Quyet (93.1).
What it will never do
A decision model generates no text, so it explains nothing. You get “0.91” and no why; if you need a justification, you pair it with an LLM. A badly framed question still returns a confident-looking score. Quality drops outside English. And Microsoft itself says it must not be the sole decision-maker for consequential calls about people: credit, employment, housing, insurance, healthcare. For those, the score is an input to a human, never the verdict. Keep your evals and your human in the loop.
In summary
| Question | Answer |
|---|---|
| What is a decision model? | A small model that picks among fixed options and returns one probability per option |
| What does it output? | JSON numbers only: no text, no explanation |
| What does it cost? | $0.042 per million input tokens; output free (there is none) |
| Where does it shine? | Routing, classification, agent guardrails, LLM-as-judge, triage |
| Where does it not belong? | Generation, open questions, and consequential decisions about people |
| The caveat | Benchmarks are vendor-reported; calibration ranks behind Jev |
The pattern to remember: keep the big LLM for the conversations and the writing, and put a signal operator in front of (and around) it for the thousand small yes/no calls your system makes every hour. One pulls levers, the other writes the reports. Together they run a rather smooth station: it’s not rocket science.
Sources: Microsoft, introducing Microsoft-Decision-1 · Microsoft Foundry catalog · eesel, Microsoft-Decision-1 explained · OrcaRouter, a model that answers with a number.
