Microsoft-Decision-1: when AI stops writing and starts choosing

Microsoft announced this week Microsoft-Decision-1, an artificial intelligence model designed to do one thing: make decisions. No fluent text generation, no open-ended reasoning — just classification, routing, verification, and scoring of predefined options, all in milliseconds. The proposition is ambitious and reveals a mindset shift that had been brewing: the idea that you shouldn't use a heavy generative model for every step in an AI pipeline.

The model is available in public preview on Microsoft Foundry and will also be available through OpenRouter. Built on top of Qwen3.5-9B — an open-weight model from Chinese AI lab Qwen (Alibaba Cloud) — it underwent post-training focused exclusively on probabilistic scoring. Microsoft promises the next version will be rebased on its own models, Microsoft AI (MAI), and even OpenAI models, signaling a long-term strategy independent of any foreign base.

What is a "decision model"

The category is new, and even Microsoft acknowledges that. In a blog post, Achint Srivastava, VP of Software Engineering at Microsoft's Office of the CTO, described decision models as a distinct class of AI systems, positioned between generative LLMs and rigid programmatic rules. They don't generate text; they return, for each possible option, a calibrated probability — a number that software consumes directly.

The mapped use cases are practical and repetitive:

- Support routing: classify tickets, select the right team, and score priority - Intent analysis: identify the intent behind a user query - Data labeling: assign consistent labels to posts, feedback, and other data - AI judgment: evaluate whether an AI-generated response meets quality criteria - Safety verification: detect jailbreak attempts and prompt injection without blocking innocent requests

The technical advantage is simple but powerful: instead of asking a large LLM like GPT-6 to solve a classification problem with prompt engineering — which might cost cents per call and take seconds — the decision model responds in milliseconds with a JSON containing calibrated probabilities. The latency and cost savings are what drive adoption.

Benchmarking: speed and accuracy

Microsoft measured Decision-1 across 36 benchmarks, spanning nearly 150,000 questions, with data kept secret during training. The result: the highest accuracy recorded. In speed, it was 2.5 times faster than H2O-Lightning-4B v1.1 (the runner-up) and 35 times faster than GPT-6 Sol (OpenAI's generative model running in decision mode).

But the story gets more nuanced when you look at independent benchmarks. H2O-Lightning-4B, for instance, leads the JevBench Composite Score among 104 open-weight systems, with a calibration score of 90, speed of 92.6, intelligence of 60, and classification of 60. Microsoft did not publish JevBench raw data — only its internal benchmark — which raises the usual question: whoever controls the measurement controls the narrative.

What's undeniable is that the entire category is growing rapidly. In less than a month, OpenAI launched its Decisions API in public beta, TypeSafe introduced Jev (a general-purpose decision model), and H2O.ai has Lightning-4B. Curiously, all of them are built on Qwen bases — which raises questions about how "innovative" the underlying technology really is.

Why this matters: the modular architecture of AI

The idea of using different models for different tasks isn't new — microservices have been doing this for decades. What's new is applying this philosophy to AI architecture. Instead of one massive LLM doing everything (generating text, classifying, deciding, routing, verifying), Microsoft proposes an ecosystem where generation, reasoning, and decision are complementary blocks, each chosen for what it does best.

This fits perfectly with the wave of autonomous agents. An AI agent that interacts with real-world systems (charge a payment, forward a ticket, approve a loan) can't rely solely on an LLM for every decision — the cost would be prohibitive and the latency unacceptable. A decision model that costs fractions of a cent per call and returns in milliseconds is the perfect component for that control layer.

Microsoft is already testing Decision-1 internally across five scenarios:

1. Scientific discovery: data labeling, AI judgment, and intent analysis 2. Search relevance: evaluating relevance of search results 3. Content classification: categorization by topic or policy 4. Code verification: evaluation against defined criteria 5. Incident routing: automatic triage by severity

The Chinese model paradox

One interesting detail: Decision-1 is built on Qwen3.5-9B, from Alibaba Cloud, the Chinese company. Microsoft says it will rebase in the coming months on MAI and OpenAI, but the reason for this decision wasn't disclosed. Possible motivations: Qwen is open-weight (zero licensing cost), already has strong classification performance, and enables rapid prototyping. But dependence on a Chinese model in a corporate AI product also raises geopolitical concerns — even if Microsoft claims the training data is "subject to Microsoft's Open Data process".

What to watch

The "decision models" category will likely expand in the coming months. The question isn't whether it will grow — the market already shows it will — but who will set the standard. OpenAI with the Decisions API? Microsoft with Decision-1? Or open-source models like H2O-Lightning? And more importantly: will this new class of models really eliminate decision costs, or just shift them to another layer of the stack?

Sources: The Register, The Verge, Microsoft Command Line

✓ Independent sources cross-checked and verified before publishing