Microsoft's Decision-1 is a 9B model built on Qwen that only scores answer options

Microsoft used an October 9, 2026 announcement to launch a language model that produces no prose. According to coverage on Shattered.io, Microsoft-Decision-1 does not chat, summarize, or draft: it reads supplied content alongside a fixed list of answer options and returns a calibrated probability for each option in a single forward pass. The model is post-trained from Alibaba's Qwen3.5-9B and shipped on Microsoft Foundry at launch, with OpenRouter availability following shortly after.

What the model does

Microsoft framed the model around four core jobs — routing, classification, prioritization, and verification — plus what it calls workflow control. The typical examples: a support ticket sent to the right team, a yes-or-no moderation call on user content, or a pass/fail check on another AI agent's output before it is allowed to proceed.

The contrast with general-purpose models is structural. A chat model generates free text that a downstream system must parse, and confidence has to be inferred from wording. A decision-scoring model returns a probability distribution over options the developer defined, eliminating the parsing step entirely. These are high-volume, low-latency decisions, and Shattered.io argues a 9-billion-parameter specialist can handle them far more cheaply than a full-sized reasoning model.

Built on a rival's weights

The detail that drew the most commentary is the foundation. The Register noted that Microsoft leaned on a Chinese lab's open-weight release to get the product out the door quickly. In its launch post, Microsoft said it post-trained Qwen3.5-9B for fast, single-pass decision scoring and will soon rebase the model on other foundations, including Microsoft AI (MAI) and OpenAI models.

Shattered.io calls that a notable admission for a company with its own frontier-model ambitions: a flagship-named product shipping on a competitor's weights, with customers told up front that the underlying engine will change later.

The numbers, and their limits

Microsoft's headline claim is the highest accuracy across a comparison spanning 36 benchmarks and roughly 150,000 held-out questions, with 83.5 percent accuracy overall — a figure also relayed by the German outlet BornCity.

Speed claims follow the same pattern. AI Weekly and TestingCatalog both report Microsoft's claim that Decision-1 runs 4.5 times faster than a model called Quyet-1.0-Large and 35 times faster than GPT-6 Sol, and Data Studios led its write-up with the 35x figure. Microsoft's own materials describe that result as approximately 35x lower P50 latency on its own benchmark — a narrower claim than a general speed win, since Decision-1 scores fixed options while GPT-6 Sol generates text. Microsoft also claims a 2.5x advantage over the runner-up in its comparison, H2O-Lightning-4B v1.1. Taiwan's Liberty Times reported response times in the neighborhood of 85 milliseconds, without specifying the hardware or test conditions behind that figure.

All of these are vendor-reported numbers. The coverage reviewed found no independent, third-party reproduction of the benchmark comparison, the speed multipliers, or the latency claim, so they should be read as launch claims rather than verified results.

One less-noticed figure may matter more in production: Microsoft says decisions stayed stable under small input changes 98.7 percent of the time, per AI Weekly's citation of the company's launch summary. Stability under trivial rewording is critical when downstream systems act on a model's calls automatically, so this number, if it survives outside testing, is a genuinely useful data point for engineering teams.

Loose ends

Microsoft is already pointing to internal uses ranging from incident response to scientific discovery. Pricing is murkier: a specific per-token figure circulated in headlines, but the coverage itself states that neither Microsoft's announcement nor Foundry documentation confirmed a price in the materials reviewed.

Why it matters

Decision-1 is a bet that the next wave of AI infrastructure is not chattier assistants but small, narrow models embedded in pipelines — routing, gating, and verifying at volume, where latency and cost per decision dominate. It also signals that open-weight bases like Qwen are now scaffolding even for the largest vendors' branded products, with the base model treated as swappable. For engineering teams, the open questions are practical: whether the accuracy and stability claims hold up under independent testing, and what it means for production systems when Microsoft follows through on rebasing the model onto entirely different foundations.