Microsoft Releases Ultra-Cheap AI Decision Model! Based on Qwen 9B, 35x Faster Than GPT-6

Just a couple of days ago, the engineering team at Microsoft’s CTO office suddenly dropped a bombshell: the official release of a scoring model specifically designed for rapid decision-making, Microsoft-Decision-1

This time, Microsoft didn’t bother competing on parameter scale for general chat and code generation. Instead, it went extremely vertical and sharp: this is a Decision Model built specifically for classification, scoring, routing, filtering, and flow control

Breaking down its technical details, there are several key points that should make every developer’s eyes light up:

1. First, the base model: It directly adopted Alibaba Tongyi’s open-source Qwen3.5-9B, which Microsoft turned into an ultra-fast decision-making powerhouse through targeted post-training fine-tuning, enabling it to output well-calibrated probability scores directly in a single forward pass
2. Second, the price: Cheap to the point of being almost free: input tokens cost only $0.042 per million, and even more remarkable, output tokens are completely free, with overall cost cut by a factor of 200 compared to traditional flagship large models
3. Third, inference speed: Absolutely outrageous: in Microsoft’s own benchmark tests, its P50 latency is a full 35 times faster than the frontier large model GPT-6 Sol

From OpenAI’s Decisions API a couple of days ago, to Liquid AI’s d1, to Microsoft’s high-profile entry today with Decision-1, an incredibly clear technical evolution path has emerged: hand heavy reasoning to large models, and high-frequency decisions to specialized small models

Introduction

What exactly is this so-called Decision Model

Over the past year or two, when building Agents or enterprise-grade AI workflows, no matter how trivial the judgment call, people would directly call GPT-6 or some other frontier large model

For example, determining whether a user’s sentence is small talk or a question? Call the large model once; deciding which external Skill to invoke? Call the large model again; evaluating whether a generated response is compliant? Call the large model yet again

This approach not only keeps end-to-end latency stubbornly high, but every single step is also burning through API budgets at an alarming rate

What’s worse is the “randomness” and “verbosity” of large models: you just want it to give you option A or B, but it insists on first analyzing three sentences of background context, and occasionally even throws a format parsing error

Microsoft Decision-1’s design philosophy completely changes this situation: it abandons lengthy autoregressive text generation and instead outputs a precisely calibrated probability distribution for a fixed set of candidate options

Core product features:

• Base architecture: Post-trained and fine-tuned based on Qwen3.5-9B, with plans to expand to Microsoft’s own MAI series and OpenAI base models in the future
• Output format: Single-pass inference, supporting Yes/No binary classification, single/multiple-choice scoring, rating, and rigorous rubric-based judging
• Access channels: Already fully available on Microsoft Foundry (Azure AI catalog) and OpenRouter
• Extremely cheap: $0.042 / 1M tokens for input, completely free for output

To help everyone understand the ecological niche of this type of specialized small model within complex systems, I’ve drawn the following next-generation dual-track Agent architecture diagram:

Why large models can’t handle high-frequency decisions well

Many people might ask: frontier large models have such high intelligence, how hard can it be for them to make a simple judgment call

In actual engineering deployment, general-purpose generative large models have three fatal flaws:

▸ 1. Latency snowballing

In a complex Agent chain involving 20 intermediate judgments, even if each judgment only adds 100 milliseconds of extra delay, the entire workflow will accumulate an additional 2 seconds of latency out of thin air; not to mention that regular LLMs often need hundreds of milliseconds or even several seconds for each round of thinking and generation

▸ 2. Fragile generalization and poor resistance to perturbation

In production environments, user instructions often come in all sorts of unpredictable forms. Just by swapping in a synonym, reordering the options, or introducing a tiny bit of formatting noise, a traditional large model’s judgment result can flip in bizarre ways

▸ 3. Difficult to calibrate probabilities

Many business decisions rely on strict confidence levels, such as auto-executing only when confidence exceeds 90%, and routing to manual review when below the threshold; the probability values output by generative large models are often severely distorted, making them difficult to use directly as engineering thresholds

Decision-1 has been extremely optimized across all these dimensions: when facing 8 types of high-difficulty input perturbations (including shuffled options, synonym rewrites, formatting noise, etc.), the average decision flip rate is only 1.3%, and in option reordering and permutation tests, it even achieved a stunning zero flip rate

Official benchmarks and comparisons

The test matrix Microsoft published this time is quite massive, covering 36 benchmark tests with nearly 150,000 evaluation questions (including blind tests on public and private datasets not involved in training)

The chart below shows the comprehensive capability scores and median latency distribution of various decision models on the authoritative industry evaluation platform JevBench:

Combined with the data officially disclosed by Microsoft, let’s look at a few core comparison metrics:

Of course, I must honestly remind everyone here: all of this data comes from Microsoft’s own engineering report. Although they claim the benchmark data was kept strictly confidential and blind-tested for model training purposes, this is not yet equivalent to a comprehensive reproduction by an independent third party, and actual deployment results still need to be verified and leveled out in each team’s own real business scenarios

Four major real-world combat scenarios inside Microsoft

Before officially going public, Microsoft had already intensively validated this internally across its production lines:

▸ 1. XBOX user real feedback and massive data annotation

The XBOX team used it to categorize specific topics from over 10,000 open-ended player comments sourced from surveys, the STEAM community, and social media. Results showed that its annotation quality was basically on par with GPT-6 Sol, but it ran over 14 times faster and cut costs by 200x

▸ 2. Copilot real-time quality control and session arbitration

When monitoring the quality of Agent responses, the Copilot R&D team replaced the cumbersome large-model judge with Decision-1, directly matching the evaluation quality of GPT-5.6 Luna, while response speed shot up by a full 100 times

▸ 3. Real-time incident response and knowledge retrieval for operations engineers

In Microsoft’s online operations alert system, faced with a complex tangle of logs, tickets, and call chains, Decision-1 took on the role of knowledge retrieval routing, demonstrating retrieval speed and accuracy that surpassed conventional LLM approaches

▸ 4. Adaptive experiment re-planning in scientific exploration

In the scientific research agent Microsoft Discovery, the agent needs to continuously review experiments and adjust strategies based on a rubric; after adopting Decision-1, scoring consistency improved by a factor of 46 compared to general-purpose large models, and overall adaptive planning speed increased by nearly 4 times

Typical deployment architecture: how to build an enterprise-grade Agent

If your team is building an enterprise knowledge base or automated agent, I strongly recommend referencing this dual-track architecture:

Complex user request input
       ↓
[Decision Model] Intent classification and difficulty assessment (filtering meaningless requests, determining whether it matches the standard knowledge base)
       ↓
[Decision Model] Model routing and Skill matching (locating the target skill from hundreds of tools in milliseconds)
       ↓
[Primary reasoning large model] Complex reasoning and actual tool chain execution (focused on high-intelligence, high-value logical computation)
       ↓
[Decision Model] Result quality inspection and safety compliance guardrail (scoring based on business rubrics, blocking jailbreak attacks)
       ↓
Secure delivery endpoint / triggering secondary confirmation by human experts

Complex user request input
       ↓
[Decision Model] Intent classification and difficulty assessment (filtering meaningless requests, determining whether it matches the standard knowledge base)
       ↓
[Decision Model] Model routing and Skill matching (locating the target skill from hundreds of tools in milliseconds)
       ↓
[Primary reasoning large model] Complex reasoning and actual tool chain execution (focused on high-intelligence, high-value logical computation)
       ↓
[Decision Model] Result quality inspection and safety compliance guardrail (scoring based on business rubrics, blocking jailbreak attacks)
       ↓
Secure delivery endpoint / triggering secondary confirmation by human experts

Under this division of labor, high-frequency and trivial gating logic is entirely handled by the specialized 9B model within milliseconds, while the primary large model can focus all its computing power on critical reasoning tasks

Overall system throughput doubles, response latency plummets, and the monthly bill can be slashed by an order of magnitude directly

Summary

The arrival of Microsoft Decision-1 sends an extremely clear signal:

The era of relying on a single trillion-parameter large model to handle everything end-to-end is coming to an end, and system-level decoupling and collaborative division of labor are the true endgame of industrial-grade AI applications

Using an ultra-fast small model with single forward-pass inference to handle classification and routing, while using heavyweight frontier models to handle complex thinking — this is the real solution for Agent deployment that truly runs fast, stays under control, and remains affordable

The model is now live on both Microsoft Foundry and OpenRouter. If you have an agent pipeline project on hand, I strongly recommend giving it a real test run

Project link: https://commandline.microsoft.com/microsoft-decision-1-model-foundry/