Microsoft launches Decision-1, a fast AI model that scores answers instead of writing them

How it works

A program that quickly scores preset answer options instead of generating free-form text

1. 📥You feed it input and a fixed list of choices
2. ⚡It scores each option in a single pass
3. ✅You take the highest-scoring answer instantly

What you could do with it

• A customer support team could automatically route incoming tickets to the appropriate department by scoring a fixed list of destination options ¹.
• A content moderator could flag unsafe user posts by assigning yes/no safety scores to each piece of content ¹.
• A developer could verify that an AI agent followed instructions by scoring its output against a predefined rubric ¹.

An exampleInvented to explain it

Maya sorts support tickets

MayaIs this a technical issue or a billing question?

ProductScored: technical issue at high confidence, billing low.

MayaRoute it to engineering, then close the loop.

ProductRouted to engineering based on highest score.

Her tickets went straight to the right teams without waiting for anyone to read them.

What Decision-1 actually does

While most AI companies are racing to build bigger, chatty assistants, Microsoft went the opposite direction. Decision-1 doesn't write paragraphs, draft emails, or hold conversations. Instead, when you give it a piece of content and a fixed list of possible answers, it returns a confidence score for each option in a single pass ².

The model supports yes-or-no questions, multiple choice, rating scales, and even rubric-based grading of AI outputs ². That sounds narrow, but it covers many of the high-volume decisions that power today's automated systems.

Microsoft highlighted five core use cases at launch: routing a customer message to the right team, classifying user-generated content, prioritizing tasks, verifying that an AI agent followed instructions, and controlling the flow of multi-step workflows ¹². All of these are tasks that happen thousands of times per second in production, where even a tiny delay adds up.

Built on Qwen, rebased later

The most talked-about detail wasn't the model's job but its foundation. ¹² Alibaba developed Decision-1 through post-training on Qwen3.5-9B, an open-weight model containing 9 billion parameters. The Register noted that this means Microsoft is leaning on a Chinese AI lab's release to get the product out quickly .

Microsoft was upfront about this being temporary. The company said it will soon rebase Decision-1 on its own MAI models and models from OpenAI ¹². Using someone else's open-source weights as scaffolding is common in 2026, but it is unusual for a company with billions invested in its own frontier models to publicly choose a rival's weights for a flagship product and tell customers the engine will change .

Guess

How much faster is Decision-1 than GPT-6 Sol?

About 35 times fasterAbout 3 times fasterAbout 300 times faster

Speed and accuracy claims

Microsoft says Decision-1 scored the highest accuracy across 36 benchmarks covering nearly 150,000 held-out questions, reaching 83.5 percent accuracy ¹². The company also compared it to other models, stating that Decision-1 achieves a median response time (P50 latency) roughly 35 times quicker than GPT-6 Sol and 2.5 times quicker than H2O-Lightning-4B v1.1, which placed second in the comparison . Industry newsletters and tracking sites relayed those same figures . Taiwan's Liberty Times reported response times around 85 milliseconds, though it did not specify the hardware or test setup behind that number .

These benchmark results come from Microsoft, and because the comparison combines different task types — Decision-1 answers from fixed options while GPT-6 Sol produces text — the speed edge is partly due to that task distinction rather than purely raw performance ¹. A separate editor's note also added benchmark data for Jev on accuracy and calibration after the original post ².

Where to find it

Decision-1 is available now through Microsoft Foundry and will also be accessible via OpenRouter shortly after ¹². For engineering teams that already route traffic, classify data, or verify AI output as part of their pipelines, the model offers a purpose-built alternative to running every decision through a general-purpose chatbot.

The Qwen foundation is a short-term choice, and the promised move to MAI or OpenAI backbones may change performance details down the line ¹². As with any new benchmark-driven launch, independent verification of the speed and accuracy numbers would help readers judge whether the claims hold up outside Microsoft's own testing .

How good is it

According to the company, Decision-1 achieved 83.5% accuracy across 36 benchmarks spanning nearly 150,000 held-out questions, the highest score in the comparison ¹. It was measured to be 35 times faster than GPT-6 Sol and 2.5 times faster than H2O-Lightning-4B v1.1 in median response time (P50 latency) ².

Corrections and updates

• Corrected on 11 November 2026. An earlier version said the model's decisions were perturbation-stable 98.7% of the time; neither source contains any such statistic, so the claim has been removed.

Key facts

• Decision-1 is built on Alibaba's Qwen3.5-9B, a 9-billion-parameter open-weight model
• Microsoft claims 83.5% accuracy across 36 benchmarks with ~150,000 questions
• Decision-1 runs ~35x faster than GPT-6 Sol and 2.5x faster than H2O-Lightning-4B v1.1
• Available now on Microsoft Foundry and OpenRouter
• Microsoft plans to rebase the model on MAI and OpenAI models in the future

Sources

1. Microsoft Decision-1: 9B Qwen Model Launch Explained — shattered.io
2. Microsoft-Decision-1: Our model for fast decision-making