Microsoft-Decision-1: the 9B model that decides instead of generating text, at $0.042 per million tokens

What it is

On October 9, 2026, Microsoft launched Microsoft-Decision-1, a model that doesn't generate text: it takes some content and a fixed set of options and returns a calibrated probability for each option. It's a "decision model" for routing, classification, prioritization, verification and workflow control in apps and agents. It's available in Microsoft Foundry and on OpenRouter.

Key numbers (per Microsoft)

• Base: post-trained from Qwen3.5-9B (9B parameters). Microsoft says it will rebase it on other models, including MAI and OpenAI.
• Price: $0.042 per million input tokens; output is free.
• Context: 32,768 tokens (OpenRouter).
• Speed: median latency ~35x lower than GPT-6 Sol and 2.5x lower than runner-up H2O-Lightning-4B.
• Quality: highest accuracy in a 36-benchmark comparison of ~150,000 held-out questions.
• Robustness: flips its decision on 1.3% of input perturbations (paraphrases, reordered options).

These are Microsoft's own results and haven't been independently verified.

Reported internal use cases

• Xbox Research: sorted 10,000+ feedback items and reviews into themes, comparable to GPT-6 Sol, 14x faster and 200x cheaper.
• Copilot: response-quality grading, competitive with GPT-5.6 Luna and 100x faster.
• Incident response and Microsoft Discovery (experiment replanning).

How it's used

You describe the "state" and typed questions (yes/no, choice, rating):

{"model": "my-deployment",
 "state": "Customer: I was charged twice this month.",
 "questions": {"team": {"type": "choice",
   "instructions": "Which team should handle this?",
   "criteria": {"billing": "Payment issues", "engineering": "Errors and bugs"}}}}

The response returns the chosen option and each option's probability. It doesn't work with chat SDKs: it uses its own decisions API.

Where it fits for SMEs

• Support triage: route tickets and emails to the right team.
• Model routing: decide whether a query goes to a cheap or expensive model.
• Agent guardrails: approve, block or escalate an agent's next action, using confidence as a threshold.
• Data labeling and customer feedback analysis.
• Grading AI responses against a rubric.

Cautions

1. Set confidence thresholds: below X%, send to a human.
2. Validate on your data before production.
3. Weights update continually (per OpenRouter): watch for regressions.
4. Not for generating text, summarizing or translating.
5. Check where data is processed (GlobalStandard vs DataZoneStandard deployments) if you have residency requirements.

Sources

• Microsoft announcement (Oct 9, 2026): https://commandline.microsoft.com/microsoft-decision-1-model-foundry/
• Microsoft Learn: https://learn.microsoft.com/en-us/azure/foundry/foundry-models/how-to/use-foundry-models-microsoft-decision
• OpenRouter: https://openrouter.ai/microsoft/microsoft-decision-1

---

Want to use AI for low-cost triage and automation? KGA IT offers a free consultation: kga-it.com.