Microsoft has rolled out a compact model that scores choices rather than chats, and the design hints at where a chatbot's next move might be heading. Microsoft just put a decision-scoring model on the shelf. The pitch is simple: feed it a situation and a fixed list of possible answers, and it returns a calibrated probability for each one.
The release, reported by Help Net Security, lands as vendors race to make AI agents less reckless. A scoring model that knows when to punt, route or escalate is a useful choke point for any automated system that touches inboxes, tickets or user data.
Quick take
- The model, called Microsoft-Decision-1, is built for routing, classification, prioritisation, verification and workflow control, according to Achint Srivastava, VP of Software Engineering in Microsoft's Office of the CTO.
- Microsoft tested it against 36 benchmarks covering roughly 150,000 questions, and said it was about 2.5 times faster than the runner-up on speed, and roughly 35 times faster than GPT-6 Sol on median latency.
- It was post-trained from Qwen3.5-9B and runs single-pass scoring, with future versions planned on other models including those from Microsoft AI and OpenAI.
- Pricing is $0.042 per million input tokens, with no charge for output tokens, and access is via Microsoft Foundry on Azure or through OpenRouter.
- It is not a general chatbot. It scores and grades; it does not write paragraphs, hold a conversation, translate or summarise.
How Microsoft tested it
The model scores yes/no, multiple-choice, rating-scale and rubric-style prompts, and the call goes through a structured API rather than a chat box. Output is a probability per option, not a free-text reply. Developers can also set thresholds so the system defers or asks for human review when confidence is low.
Internal pilots give a feel for the workload. The Xbox team fed it more than 10,000 pieces of survey, store and social feedback, and Microsoft reported results competitive with GPT-6 Sol at more than 14 times the speed and around one two-hundredth of the cost. The Copilot team used it to grade agent replies, reporting competitive accuracy at roughly 100 times the speed. Inside Microsoft Discovery, evaluation scores were 46 times more consistent than the LLM-based baseline, and the adaptive replanning loop ran nearly four times faster overall.
Why this matters
Decision-scoring is plumbing, not poetry, and plumbing is what most enterprise security stacks still lack. A small, cheap scorer that can rank options, grade outputs or veto a dodgy answer is the kind of layer that lets a larger chatbot take a back seat. For defenders, that means cheaper, faster triage of suspect mail, sharper prioritisation of alerts, and a way to grade the output of other AI tools without sending every prompt to a frontier model.
The trade-off is obvious. Calibration is only as good as the rubric, and rubrics are written by humans who miss edge cases. Microsoft itself flags speed, robustness, calibration and safety as challenges. Treat the scores as guidance, not gospel, and tune the thresholds against your own data before letting the model act on its own.
A worked example: routing suspicious mail
Say you run a security team and your inbox triage bot gets 5,000 forwarded "is this phishing?" reports a week. Today, every report gets dumped into a queue and a human analyst sorts it. With the decision model wired in as a front gate, the bot scores each message against fixed buckets: known campaign, credential phish, vendor impersonation, marketing, junk. High-confidence cases get auto-closed or auto-escalated with a reason code; low-confidence ones go to a human. The scorer is cheap, so the cost of being wrong on the first pass is small. The win is that your analysts stop being a sorting function and start being an investigation function.
| Role | Old workflow | With the scoring model |
|---|---|---|
| Triage analyst | Sorts every report by hand | Reviews only low-confidence cases |
| Cost per decision | Sends everything to a large model | Sends only the hard ones |
| Audit trail | Ad-hoc notes | Reason code per closed case |
What to watch next
Keep an eye on three things. First, whether Microsoft ships the promised versions based on its own and OpenAI's models, and whether the calibration holds up outside Microsoft's own benchmarks. Second, how rivals respond with similar scorers, since OpenRouter already routes multiple providers and pricing pressure is the obvious lever. Third, whether security teams actually adopt these models as gates or let them slip into autonomous roles without thresholds. A scoring model is a brake pedal. Use it like one.
Key takeaways
- A cheap, fast decision-scoring model is more useful for security teams than another chatty chatbot.
- Probability outputs only help if you set confidence thresholds and review the rubrics behind them.
- Watch pricing, calibration claims and whether vendors let these models act without a human in the loop.
Frequently asked questions
What is Microsoft-Decision-1 in one line?
A small model that scores a fixed list of possible answers and returns a calibrated probability for each, built for routing and grading tasks rather than free-form chat.
Can it replace a large language model in a security stack?
No. Microsoft describes it as a scoring and grading layer, and the company itself says it is not designed for open-ended generation, conversation, translation or summarisation.
How much does it cost to run?
Microsoft is charging roughly four cents per million input tokens, with no fee on the output side, based on the figures reported by Help Net Security.
Where can developers try it?
It is generally available through Microsoft Foundry on Azure and through OpenRouter, which routes multiple model providers through one API.
What should a security team validate before deploying it?
The rubric, the confidence thresholds, the fallback path for low-confidence scores, and how the system logs overrides for audit.