The Pulse
Microsoft Launches Decision-1, a Model for Fast Structured Choices
Microsoft released Decision-1, a model that scores predefined choices for tasks such as classification, routing and workflow control. The company says it led a 36-benchmark comparison and costs $0.042 per million input tokens, with output t
AI.info Team · Written with AI ·
“Microsoft-Decision-1 achieved the highest accuracy in our 36-benchmark comparison, spanning nearly 150,000 questions across benchmarks kept blind from training.”
Achint Srivastava, VP of Software Engineering, Office of the CTO, Microsoft
Microsoft announced Decision-1 on October 9 as a model for applications that need to choose among fixed options, rather than generate open-ended text. The company says the model ranked first for accuracy in its comparison and ran faster than the other systems it tested. Developers can use it through Microsoft Foundry and OpenRouter.
Decision scores, not generated answers
Decision-1 starts with a defined set of possible answers and returns a calibrated probability score for each option. It supports yes-or-no questions, multiple choice, rating scales and rubric-based grading of AI responses or agent actions. Applications can use the output to classify a request, route it to a tool or team, verify a condition, or decide whether an agent should continue or ask for review.
Microsoft says it built the model by post-training Qwen3.5-9B for fast, single-pass decision scoring. The company plans to rebase future versions on other models, including models from Microsoft AI and OpenAI. That distinction matters for developers: Decision-1 is aimed at narrow, repeatable choices, not tasks that call for extended reasoning or free-form writing.
Microsoft’s speed and accuracy claims
The company says Decision-1 ranked highest for accuracy across 36 benchmarks and nearly 150,000 questions that were kept blind from training. Microsoft reports it ran 2.5 times faster than H2O-Lightning-4B v1.1, the runner-up on speed, and 35 times faster than GPT-6 Sol. Those are Microsoft’s benchmark results; the announcement does not describe an independent evaluation.
Microsoft also tested how often small changes to the same request altered the model’s choice. Across eight kinds of perturbation, it says the decision changed 1.3% of the time on average, with no flips when options were reordered or their descriptions paraphrased. For safety, the company tested 5,250 requests across 11 benchmarks covering harmful content, jailbreak attempts and prompt injection, and says the model refused harmful requests while retaining utility.
Internal tests focus on repeated work
Microsoft cites several uses inside the company. Xbox Research used Decision-1 to sort more than 10,000 open-ended survey and review comments into themes defined by researchers; Microsoft says the model matched GPT-6 Sol on quality while running more than 14 times faster and costing 200 times less. The post also says the Copilot team found it competitive with GPT5.6 Luna for response-quality checks while running 100 times faster.
For Microsoft Discovery, the model graded experiments in an adaptive-replanning process. Microsoft says Decision-1 was 46 times more consistent than an LLM-based score at three times the speed, and that the replanning process ran nearly four times faster. The announcement gives no underlying timing figures for those internal comparisons, so the relative results do not show how long a full task takes.
What developers pay
Microsoft lists input tokens at $0.042 per million and output tokens as free. The company presents the price and speed as reasons to use a focused decision model for frequent steps in an application instead of sending each one to a general-purpose language model. The economics will depend on the task, the accuracy required and the amount of traffic.
Decision-1’s central promise is narrower than a general assistant’s: return a structured choice that software can act on. That makes validation on a developer’s own examples especially important, since an incorrect classification or route can still trigger an automated action even when the answer arrives quickly.