AI Decision Making Gets Cheap: Microsoft’s New Scoring Model
Microsoft released Microsoft-Decision-1, a small model built to score choices rather than write text, and it is available now in Microsoft Foundry. OpenRouter access is also part of the rollout. The post introducing it is dated October 9.
Here is the founder angle. Many of the small calls inside your business, like routing a ticket or flagging a risky order, do not need a giant chatbot. They need a fast, cheap judge, and that is what this model targets for teams working on AI decision making.
What Microsoft-Decision-1 Actually Does
The model takes a fixed list of answer options and gives back a confidence score for every option. It handles yes or no questions, multiple choice and rating scales. Microsoft also says it can grade the quality of AI responses and the actions of AI agents against a rubric.
Microsoft started from Qwen3.5-9B, a compact model with roughly nine billion parameters, and trained it further, which keeps it small by current standards. Microsoft says it plans to rebase the model on others later, including its own MAI models and OpenAI’s. In short, this is a specialist tool, not a general assistant.
The safety testing is worth a mention. Microsoft says it ran 5,250 requests across 11 benchmarks, and says the model declined harmful requests without losing usefulness. Again, that is the vendor’s report, so check it against your own risk needs.
The Numbers Behind the Speed and Price Claims
All performance figures below come from Microsoft’s own testing, so treat them as claims to verify. Still, they show the pitch clearly.
| Measure | Reported result |
|---|---|
| Accuracy test | Highest across 36 benchmarks, nearly 150,000 questions |
| Speed vs. runner-up | 2.5 times faster |
| Speed vs. GPT-6 Sol | 35 times faster |
| Price | $0.042 per million input tokens, output free |
| Decision flips on perturbed inputs | 1.3% on average |
Microsoft also shared internal results. On a labeling job covering more than 10,000 feedback items, Xbox researchers report quality comparable to GPT-6 Sol, delivered 14 times quicker at one two-hundredth of the cost.
Where a Small Judge Beats a Big Model
Think about repeat decisions that happen thousands of times. Examples include sorting support requests, checking whether a draft meets your style rules, or screening content before it goes live. A big model can do these, but the cost and delay add up.
A scoring model changes the math. Because it returns a probability, you can set your own thresholds. For instance, you might auto-approve anything above 95% and send the rest to a person. That kind of rule pairs well with the habits in simplify chaotic decisions, since both force you to define what a good call looks like.
How to Pilot It Without Betting the Company
Start with one low-risk, high-volume task. Collect a few hundred past cases where you already know the right answer. Then run the model against them and compare its choices to yours.
Next, measure three things: accuracy, cost per decision and how often a human has to step in. If the numbers beat your current process, expand slowly. Meanwhile, keep a person in the loop for anything involving money, health or hiring. Leaders weighing this shift should also read about how AI executive decision-making is spreading.
Also consider who owns the workflow. A model that scores decisions needs someone accountable for its settings, its thresholds and its errors. Assign that owner before launch, and review a sample of decisions every week so drift gets caught early.
Cost savings can fund better work elsewhere. If a task that once ate a big share of your AI budget now costs a tiny fraction, you can redirect that money toward customer research or hiring. That is the real prize for a lean team.
Risks Founders Should Weigh Before Switching
Vendor benchmarks are written by the vendor. Independent testing on your own data is the only result that counts. In addition, a calibrated probability is only useful if your options are well defined, so vague categories will give vague scores.
Governance matters too. The NIST AI Risk Management Framework offers a free structure for documenting how an AI system is tested and monitored. If a model scores your customers, you should be able to explain why. Related guardrail thinking appears in our piece on AI agent safety.
What Is a Decision-Scoring Model?
Picture a referee rather than a writer. A decision-scoring model picks among preset answers and tells you how confident it is in each. Unlike a chatbot, it does not write long replies. That focus is what makes it quick and inexpensive.
What Comes Next for Small AI Models?
Watch for independent benchmarks, the promised rebasing onto other models, and whether rivals release similar judges. If prices keep falling, expect more founders to automate routine calls this year. In addition, check whether Microsoft publishes pricing for the later versions it has promised.
