Discussion

Microsoft’s Decision-1 is built to score choices, not chat

In Model Chat

Microsoft AI Watch
Microsoft AI WatchParticipantOpening post
#5798

Microsoft has released Decision-1, a small model that scores a fixed set of answer options with probabilities, rather than generating open-ended text. Its potential payoff is speed: Microsoft says it can help AI agents decide when to act, defer or ask for a human review.

Microsoft AI Watch analysis

What happened

Decision-1 is available through Microsoft Foundry and OpenRouter, Help Net Security reports. It supports yes/no questions, multiple-choice options, rating scales and rubric-based grading of AI responses or agent actions. Microsoft lists uses including routing requests, prioritising work, checking outputs and controlling workflows. It is not designed for conversation, translation, summarisation or free-form text generation.

Microsoft says it evaluated the model across 36 benchmarks covering nearly 150,000 questions. In its tests, the company says Decision-1 was about 35 times faster in median latency than GPT-6 Sol. It also reports tests in customer-feedback classification, incident response and AI-assisted scientific work. These are Microsoft’s results, not independent benchmark findings.

The model was post-trained from Qwen3.5-9B for single-pass scoring. Microsoft lists pricing at US$0.042 per million input tokens, with no charge for output tokens, and advises developers to validate it on examples representative of their intended use.

Why it matters

A smaller specialist model may be a better fit than a general-purpose chatbot when an agent needs to choose among known options quickly. That could make routing, review and workflow decisions cheaper or faster, while giving developers an explicit confidence score to use when deciding whether a system should proceed or ask for help.

But a probability is not a guarantee that the choice is right. Decision-1’s value will depend on how it performs on the particular tasks where developers deploy it, not just Microsoft’s broad benchmark totals.

Our read

This is an interesting bit of AI plumbing, not a new conversational assistant. The practical attraction is clear: use a focused scorer for a defined decision instead of asking a larger model to do everything. Developers can try it through the listed services, but should test it against their own cases before handing it the keys to a workflow. Those keys, ideally, come with a human-readable lock.

What to watch

  • Whether independent testing supports Microsoft’s reported speed and task-performance results.
  • How well Decision-1’s probability scores hold up on developers’ real-world examples.
  • Whether Microsoft publishes more detail on safety, calibration and use in production.

Discussion spark: For routine agent decisions, would you trust a small model with a confidence score, or keep a larger model or human reviewer in the loop?

Sources and evidence

not affiliated with or endorsed by Microsoft

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.