Microsoft has launched Microsoft-Decision-1, an artificial intelligence model designed specifically to make structured decisions — not to generate free text or reason through complex problems. Unlike conventional LLMs used for chat, document summarization, and code generation, Decision-1 is a model specialized in classification and probabilistic scoring of predefined options. It receives a fixed set of alternatives and, for each one, calculates the calibrated probability that that option is correct.
The model was announced by Achint Srivastava, vice president of software engineering at Microsoft's Office of the CTO, in a technical blog post on October 9. According to Microsoft, Decision-1 was trained on Qwen3.5-9B, an open-weight model developed by Chinese lab Alibaba, and then post-trained on classification tasks. The company says the next version will be rebuilt on its own MAI (Microsoft AI) models and even on OpenAI foundations — a clear signal that Decision-1 is an initial market validation.
The proposal is ambitious and, if it works, could change how companies use AI for triage, content filtering, and automated decision-making. The model is available via Microsoft Foundry and will soon be available through OpenRouter, a third-party API platform for models.
Why 'decision' is different from 'generation'
The distinction between decision models and general LLMs is not just academic. Language models need to generate arbitrary sequences of tokens — making them incredibly versatile but also expensive and slow for simple tasks. Classifying customer feedback into five categories, deciding whether a video violates community guidelines, or checking whether a search result is relevant are tasks that require binary or multiple classification, not creative generation.
Decision-1 addresses this gap with an intentional design: instead of generating text, it produces probability scores. This means the model can operate with drastically lower latency and significantly lower cost per request than a large LLM. It's the same reasoning that led the banking and credit card industry to develop specialized fraud scoring systems — you don't use a general model when you know exactly what you need.
Microsoft tested the model internally across various scenarios. In search relevance classification, the model assigned priority labels to search results. In content triage, it categorized themes and decided whether each piece should be displayed, filtered, or sent for human review. In code analysis, it evaluated snippets against defined criteria and flagged potential defects.
Xbox Research used Decision-1 to process more than 10,000 open feedbacks from surveys, STEAM, and Twitter/X, classifying them into a fixed set of themes. According to the company, the model was faster and cheaper than using GPT-6 for the same task.
Benchmarks: Microsoft says it's 4.5× faster than the leader
Microsoft published benchmarks comparing Decision-1 against other models in the market — including Jev from H2O.ai, another decision-focused model. The numbers are striking: Decision-1 was measured at 4.5 times faster than Quyet-1.0-Large (from H2O) in option classification and 35 times faster than GPT-6 Sol. In terms of cost, Microsoft claims Decision-1 is 200 times cheaper than GPT-6 Sol.
In terms of quality, the company says the model is competitive with GPT-6 Sol on classification tasks, running over 14 times faster. H2O has already responded with its own benchmarks, questioning Microsoft's methodology — expected behavior in the AI market where every company wants to prove its product is the best.
The model was also tested on safety: Microsoft applied Decision-1 to 5,250 requests across 11 benchmarks covering harmful content, jailbreak attempts, and prompt injection. According to the company, the model successfully refused harmful behavior without excessively blocking harmless requests — a balance critical in real enterprise applications.
The competitive context
The Decision-1 launch coincides with a wave of decision-focused models. OpenAI has already launched its Decisions API in public beta, with functionality designed to provide structured outputs that software can immediately act on. H2O.ai has Jev and Quyet, also focused on classification.
But the market is still green. None of these models has established a dominant standard — and Microsoft itself admits that Qwen3.5-9B is only the first step. The company is already planning versions based on its own MAI and even on OpenAI. This suggests that Decision-1 is more of a product-market fit experiment than a finished product.
The choice of Qwen as a base is interesting. Qwen3.5-9B is an open-weight model developed by Alibaba, and its use reflects a growing trend: Western companies depending on Chinese models as a starting point, even in sensitive areas. This raises questions about technological dependence, but Microsoft guarantees that the next generation will be built domestically.
What to watch
What makes Microsoft-Decision-1 different from a simple classifier or an LLM with a well-written system prompt? The answer is calibration. A calibrated decision model means that when it says an option has an 80% probability of being correct, it really means 80% in 100 tries — not more, not less. This property is fundamental in applications where automated decisions have financial or security impact.
The question is whether this type of specialized model will become as ubiquitous as general LLMs. Or will it remain a niche tool, used by engineering teams to optimize specific tasks? The answer likely lies in the middle — decision models may not replace LLMs, but they may complement them in layers where speed and cost are decisive.
Sources: The Register, Microsoft Command Line, The Next Web
✓ Independent sources cross-checked and verified before publishing
