Microsoft launches specialised AI model for fixed-choice decisions

Microsoft has introduced a new artificial intelligence model designed specifically for making structured decisions rather than generating open-ended text. Microsoft-Decision-1 returns calibrated probability scores for predefined options, handling tasks such as multiple-choice questions, yes-or-no queries and rating scales. This approach differs from standard generative models by focusing on classification and verification workflows instead of creative writing.

What we know

Achint Srivastava, vice-president of software engineering in Microsoft’s Office of the CTO, announced the model on 9 October. According to the Foundry model card, Microsoft post-trained the system using Alibaba’s open-weight Qwen3.5-9B base. Srivastava described decision models as an emerging AI category, noting that Microsoft plans to rebase the tool on other architectures, including OpenAI models and its own MAI series, in the future. The current version accepts only text input and supports a context window of 32,000 tokens.

Microsoft claims Microsoft-Decision-1 achieved the highest accuracy in its internal comparison across 36 benchmarks. These tests involved nearly 150,000 questions that were kept blind from the training process. The model also underwent safety evaluations on 5,250 requests across 11 benchmarks, covering harmful content, jailbreak attempts and prompt injection scenarios. Despite these safeguards, the model card explicitly advises against using the tool as the sole basis for consequential decisions affecting individuals, such as employment or credit approvals.

Performance metrics highlight significant speed advantages over larger competitors. In Microsoft’s tests, the model operated 2.5 times faster than the runner-up, H2O-Lightning-4B v1.1, and 35 times faster than OpenAI’s GPT-6 Sol. When researchers reworded or reordered identical requests, the model altered its decision in only 1.3% of cases on average, indicating stable consistency. The system also outperformed TypeSafe AI’s Jev model by a factor of 2.8 in latency tests, according to reports from The Register.

Real-world applications within Microsoft demonstrate the model’s efficiency. Xbox Research used it to categorise more than 10,000 pieces of player feedback into thematic groups. Microsoft stated this process was competitive with GPT-6 Sol in quality but 14 times faster and 200 times cheaper. Similarly, the Copilot team found the model comparable to GPT5.6 Luna in performance while operating 100 times faster. These use cases include routing, classification, prioritisation and workflow control.

Access to Microsoft-Decision-1 is available through Microsoft Foundry and OpenRouter, the latter acquired by Stripe in August. Pricing is set at $0.042 per million input tokens, with output tokens provided free of charge. The model’s rapid adoption follows the launch of Jev three weeks prior, which became the fastest-adopted model on Vercel this year. Microsoft’s entry adds another option for developers seeking cost-effective, high-speed decision-making tools.

Defender's checklist

• Verify that automated decision systems include human oversight for high-impact outcomes such as hiring or financial assessments.
• Review model documentation for explicit warnings regarding bias or limitations in consequential decision-making contexts.
• Monitor input costs and token limits to ensure budgetary control when integrating new AI services into production workflows.
• Test model consistency by rephrasing identical queries to check for stability in probability scores and final decisions.

What to do and how to stay safe: Microsoft

• Check if current classification tasks could benefit from a dedicated decision model instead of a general-purpose generative AI.
• Review vendor documentation for specific guidance on acceptable use cases and prohibited high-stakes applications.
• Compare latency and cost metrics against existing tools to assess potential efficiency gains for routing and verification tasks.
• Ensure that automated decisions are logged and auditable, regardless of the model’s speed or accuracy claims.

General security guidance from the TechWire Beat newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Can Microsoft-Decision-1 generate creative text or stories?

No, the model is designed to pick between fixed answer options and return probability scores, not to write open-ended text.

Is the model suitable for making final hiring or credit decisions?

No, Microsoft explicitly states it should not be the sole basis for consequential decisions about people.

How does the cost compare to other models?

Input costs $0.042 per million tokens with free output tokens, making it significantly cheaper than larger competitors like GPT-6 Sol.