Microsoft has introduced Microsoft-Decision-1, a model designed to return structured decisions rather than open-ended prose, with reported pricing of $0.042 per million input tokens and no charge for output tokens. For developers who need software to classify something and move on, that is a more useful proposition than another exceptionally talkative assistant. The first version is based on Alibaba Cloud’s Qwen3.5-9B, rather than a Microsoft or OpenAI model. Microsoft says that foundation will change in a forthcoming version.
Microsoft AI Watch analysis
What happened
Microsoft announced Decision-1 on 9 October. The model is offered through Microsoft Foundry, with OpenRouter availability described as coming soon, according to Thomas Claburn’s report in The Register on 10 October.
Decision models constrain answers to a limited set of responses, rated by probability. The practical distinction is that software gets an output it can act on directly, rather than a paragraph it must interpret. Microsoft’s engineering executive Achint Srivastava describes the category as purpose-built for structured outputs and low-cost classification.
Microsoft claims Decision-1 achieved 83.5% accuracy across 36 benchmarks. It also claims 2.5 times faster performance than H2O-Lightning-4B and 2.8 times faster performance than Jev in its latency test, alongside a cost advantage of more than 20 times over OpenAI’s GPT-6 Sol for text classification. Those are Microsoft’s comparisons, not a universal ranking across workloads.
Why it matters
Not every AI task needs an essay. If an application needs to select a category or choose among defined options, paying a general-purpose model to generate and explain an answer can add expense and delay without adding much value.
Decision-1 gives developers another specialised option for that work. Its pricing makes it worth evaluating where classification happens repeatedly, although the useful measure remains cost per correct decision, not simply cost per token. A cheap wrong answer can still send a workflow in the wrong direction.
The Qwen foundation is also notable: Microsoft is using an externally developed open-weight model for the initial product while signalling a move to Microsoft and OpenAI models. Customers therefore have both a new service to evaluate and a forthcoming foundation change to watch.
Our read
This is a worthwhile addition to the AI toolbox because it starts with a bounded job rather than asking one enormous model to do everything. Sometimes the clever engineering decision is to stop inviting the chatbot to every meeting.
The right next step is a comparison on representative inputs: measure classification accuracy, latency and total cost against the system already doing the job. Restricted outputs make integration simpler, but they do not make the underlying decisions infallible. Probability scores also need testing before software treats them as dependable confidence estimates.
What to watch
- Whether Decision-1’s reported price and speed advantages hold on developers’ own classification workloads.
- When OpenRouter access becomes available and under what service terms.
- What changes when Microsoft replaces the Qwen foundation, particularly for existing integrations.
Discussion spark: Would you replace a general-purpose LLM with a cheaper decision model for classification, or would the extra model dependency outweigh the savings?
Sources and evidence
- Microsoft leans on open weight model from Chinese AI lab to challenge Jev (10 October 2026, 07:10 UTC)
not affiliated with or endorsed by Microsoft