Decision Scoring, Not Generation: Microsoft's 9B Microsoft-Decision-1 and the Reported 35x Speed Edge
One forward pass, one probability per option, and no sentence of prose in return: that is the design brief behind Microsoft-Decision-1, the 9B decision-scoring model Microsoft revealed on October 9, 2026. Read as an industry data point rather than a product page, the launch cuts against the prevailing 2026 script. Where most announcements compete on scale, verbosity and context length, Microsoft spent its October statement on a system that deliberately refuses to chat, summarize or draft.
For global AI observers and enterprise teams watching inference economics, the details underneath matter more than the novelty: an Alibaba open-weight base, speed multipliers that remain vendor-reported, a pricing structure that charges almost nothing on the output side, and a workload category — high-volume, low-latency decisions — that general-purpose APIs have largely been billing at full-model rates.
Launch Facts at a Glance
- What shipped: Microsoft-Decision-1, a 9B scoring model post-trained from Alibaba's Qwen3.5-9B that reads supplied content plus a fixed list of answer options and returns a calibrated probability for each option.
- When and where: announced October 9, 2026; in public preview on Microsoft Foundry at launch, and listed on OpenRouter under microsoft/microsoft-decision-1 within roughly a day.
- Named launch tasks: routing, classification, prioritization, verification and workflow control, plus model-output evaluation described by Microsoft as AI judging and agent guardrails.
- Headline reported numbers: 83.5 percent accuracy across 36 benchmarks and approximately 150,000 held-out questions, roughly 35× lower P50 latency than GPT-6 Sol, 4.5 times faster than Quyet-1.0-Large, 2.5× faster than H2O-Lightning-4B v1.1, and perturbation-stable decisions 98.7 percent of the time.
- Reported price: $0.042 per million input tokens with output free, a figure drawn from OpenRouter's listing rather than Microsoft's own documentation.
- Standing caveat: the coverage reviewed contains no independent reproduction, so every performance claim below should be treated as vendor-reported.
What the Model Does — and What It Refuses to Do
Most of 2026's model coverage described systems producing more: more tokens, more context, more creative output. Microsoft-Decision-1 runs the other way. According to the company's announcement, it takes input content together with predefined answer options and emits a probability distribution over the options a developer supplied, instead of open-ended text. That places it in a different functional category from GPT-6 Sol, Gemini 4 Argon and Claude Opus 5.5, even though its lineage is a general-purpose language model.
The operational consequence is the elimination of parsing. A chatbot reply has to be read back, with confidence guessed and formatting instructions trusted; here the output is already numeric, ready for application code to branch on. Microsoft framed the sweet spot as high-volume, low-latency decisions: routing a support ticket to the right team, issuing a yes/no moderation call on user content, or giving an AI agent's output a pass/fail check before a pipeline advances.
That narrow scope is precisely what makes the release worth tracking. While other vendors advertise bigger, chattier assistants, Microsoft's engineering-facing pitch is a model built to pick an item from a list quickly and state confidence while doing it — which for agent pipelines, routing layers and automated verification systems may beat another general-purpose chatbot.
The Base-Model Question: Qwen3.5-9B
The detail drawing the most commentary was not the use case but the foundation. Microsoft-Decision-1 is post-trained from Qwen3.5-9B, a 9B open-weight release from Alibaba. The Register framed the launch around exactly that point — Microsoft leaning on a Chinese AI lab's open-weight model to reach market quickly — while noting the company's stated intention to migrate the base over time.
Microsoft has been explicit that the current base is a starting point rather than a destination. Its launch post states: "To build Microsoft-Decision-1, we post trained Qwen3.5-9B for fast, single-pass decision scoring and will soon rebase it on other models, including Microsoft AI (MAI) and OpenAI." For a company that has poured billions into its own model development through Microsoft Foundry and its OpenAI partnership, the wording reads as a speed-to-market choice on a narrow task, not a long-term architectural commitment.
Reusing another lab's open-weight model as scaffolding is now routine. Compact, task-specific tools follow the same logic that led JetBrains' Mellum2.1 coding model to favor a smaller parameter count over raw scale. What stands out is a vendor with frontier ambitions publicly naming a rival lab's weights inside a flagship-branded product while telling customers the engine underneath will change later.
Benchmarks, Latency and the Skepticism They Require
Microsoft justified the product with numbers, and at unusual breadth. The company says Decision-1 took the highest accuracy in a comparison spanning 36 benchmarks and about 150,000 held-out questions, a scale designed to signal breadth instead of one favorable test, and reported 83.5 percent accuracy across that evaluation — a figure the German outlet BornCity also relayed.
Speed claims arrived with the same specificity. Industry newsletter AI Weekly and AI-tracking site TestingCatalog reported Microsoft's assertion that Decision-1 runs 4.5 times faster than a model it calls Quyet-1.0-Large and 35 times faster than GPT-6 Sol; a separate write-up from Data Studios led with the 35x multiplier. Microsoft's own launch material narrowed the GPT-6 Sol result to approximately 35× lower P50 latency measured on its own benchmark, and put Decision-1 2.5× ahead of H2O-Lightning-4B v1.1, the runner-up in the comparison. That P50 framing is tighter than a general-purpose speed advantage implies, because the systems perform different tasks: Decision-1 scores fixed options while GPT-6 Sol generates text. Taiwan's Liberty Times reported response times in the neighborhood of 85 milliseconds without specifying hardware or test conditions.
All of it remains company-side. None of the coverage reviewed here included a third-party reproduction of the 36-benchmark comparison, the 4.5 times and 35 times multipliers, or the 85 milliseconds latency figure. That does not make the numbers false, but it makes them vendor-reported performance claims rather than peer-reviewed results, at least until someone outside Microsoft runs the same tests.
One less-publicized figure may matter most in production. AI Weekly cited a launch summary stating that the model's decisions were perturbation-stable 98.7 percent of the time. For any system whose binary or multi-option calls trigger automated action, stability under small input changes is a genuine engineering requirement: a routing decision that flips because of a trivial rewording is a production bug waiting to happen, so a high stability number is a useful data point for engineering teams if it survives outside testing.
Pricing Signal: Cheap Input, Free Output
Microsoft's own announcement and the Microsoft Foundry documentation reviewed here did not publish a confirmed price. What circulated came from OpenRouter's model listing and was picked up by AI Weekly and the Italian outlet Pasquale Pillitteri: $0.042 per million input tokens, with output free. A $0.042/M structure is coherent for a model whose entire output is a short list of probability scores, because there is effectively nothing to meter on the generation side.
At that level, Decision-1 is positioned as a genuinely cheap option for high-volume routing and classification — the kind of request that today often goes to smaller general-purpose models mostly because nobody wants to pay full-model prices to ask "is this spam, yes or no." The cost logic echoes Anthropic's decision to cut Claude Haiku pricing by 75 percent earlier this year: cheap, fast models win the high-frequency, low-complexity slice of API traffic even when a flagship would technically do the job as well.
Availability, Openness and Deployment Modes
Decision-1 entered public preview on Microsoft Foundry, where documentation describes it as available for classification, routing, ranking, grading and binary decisions. Microsoft's launch post initially called OpenRouter availability "coming soon," but OpenRouter's own listing confirmed the model had gone live under the identifier microsoft/microsoft-decision-1 within roughly a day of the announcement, with the platform posting that it was "live on OpenRouter."
Weights are not being distributed, according to reporting from Data Studios, so developers reach Decision-1 exclusively through hosted APIs rather than downloading and self-hosting it — a contrast with the fully open Qwen3.5-9B beneath it. That pattern is common for commercial derivatives of open-weight models, but it is worth flagging for teams that assumed a Qwen-based product would carry the base model's openness. Documented deployment options include Global Standard and, in selected regions, Data Zone Standard.
Interface Limits, Thresholds and Calibration
The model accepts text and supports inputs up to 32,000 tokens. It returns calibrated scores for predefined options, does not generate explanations, and does not handle images, audio or video — so teams needing a written rationale or multimodal input must pair it with a general-purpose model rather than expect Decision-1 to cover that work.
The confidence score is the part applications act on. The probability itself is exposed through the API rather than being merely a ranking signal, and applications use confidence to decide when to act, defer, or ask for review. One routing example pages the on-call engineer immediately when urgent is above 0.9 and priority is above 1.5. Another sets a floor: if the top probability is below 0.65, a local policy can avoid downgrading, keep the current model, or request manual selection.
A worked documentation example shows the model returning a weighted priority score of 1.9868 with 98.08% confidence. Calibration follows the standard convention teams should apply to their own data: a 90% prediction should be right about nine times out of 10 on representative cases. The quickstart needs Python 3.10 or later, so a first pilot can stay small, and because Microsoft describes pricing as dynamic, the current Foundry listing is worth checking before sizing a deployment.
A Scoring Call, End to End
The mechanics are straightforward next to prompting a chatbot. A developer supplies context plus a closed set of options, and the model returns a probability for each one. In the documentation's illustrative flow, the customer message "My payment failed twice today." was offered four options under microsoft/microsoft-decision-1:
| Option offered | Probability returned |
|---|---|
| billing_support | 0.81 |
| technical_support | 0.11 |
| spam | 0.01 |
| escalate_to_human | 0.07 |
That output format is the entire point. An application can route automatically above 0.75 and send anything lower to a human reviewer without writing a single line of text-parsing logic — a pattern familiar to anyone who has built a classification layer with traditional machine learning, repackaged on top of a modern transformer base.
Nadella's Framing and the Internal First-Mover List
Satya Nadella introduced the model personally on X, writing: "Introducing Microsoft-Decision-1, our new model for fast decision-making. It delivers top performance on structured decision tasks, outperforming both LLMs and other decision models in latency and quality. We're already testing it across Microsoft for everything from incident response and quality control to scientific discovery."
That wording names three internal arenas and signals Microsoft is not only shipping Decision-1 to third-party developers but running it through its own operational workflows first. Incident response is the logical opening move: triaging alerts, scoring severity, or deciding whether a ticket escalates to a human all fit the fixed options plus confidence score shape. Quality control and scientific discovery are broader claims, and no named Microsoft executive beyond Nadella has gone on record describing how those workflows use the model, although Microsoft-published material has since detailed other internal trials.
Documented Trials and Example Decisions
Microsoft reported that Xbox Research used Decision-1 to categorize more than 10,000 pieces of open-ended feedback into predefined themes, with quality described as competitive with GPT-5 and execution reported as 80–100 times faster. Other company-published material says teams also evaluated the model for Copilot-response assessment and adaptive replanning. These are self-reported results, so they show the intended shape of the work more than an independent measurement.
Documentation also gives a concrete picture of the decisions in mind. One example is an incident described as "The checkout service has been unavailable for 20 minutes. Customers cannot complete purchases, and there is no workaround." The model scores that report against priority levels, and the application decides what to do with the result. Among the supported uses Foundry lists:
- support-ticket triage,
- incident prioritization,
- content filtering,
- model, tool and agent selection,
- model-output evaluation.
A Microsoft developer community post on decision routing for GitHub Copilot in CLI and VS Code applies the same idea to model selection, with the router placed in the native /model command in Copilot CLI and in the Chat Model Picker in VS Code. The practical takeaway for a team evaluating the model is to start with a decision that already has clear options, such as a ticket queue or a model picker, and to measure the scores against past human decisions before letting them trigger any action automatically.
How Decision-1 Lines Up With Other Small, Specialized Models
Microsoft is not the only company betting that the 2026 AI market has room for compact, task-specific models alongside the frontier giants. Amazon shipped its own entry, Strands Decider 2B, a 2B open-source agent model built for fast decisions and clocking around 100 milliseconds in AWS's own testing. That is a smaller parameter count than Decision-1's 9B base, and open source where Microsoft-Decision-1 is API-only, but the two products chase the same niche: agent guardrails and routing decisions that do not need a full reasoning model.
| Model | Vendor | Base / size | Primary job | Output form | Price as reported |
|---|---|---|---|---|---|
| Microsoft-Decision-1 | Microsoft | 9B (Qwen3.5-9B base) | Routing, classification, verification | Calibrated probability per option | $0.042/M input tokens, free output (reported) |
| Strands Decider 2B | Amazon (AWS) | 2B, open source | Agent guardrails, fast decisions | Scored decision output | Open weight, self-hostable |
| Mellum2.1 | JetBrains | 12B | Code completion | Generated code | Bundled with IDE tooling |
| GPT-6 Sol | OpenAI | Not disclosed | General reasoning and generation | Open-ended text | Tiered ChatGPT/API pricing |
| Gemini 4 Argon | Not disclosed | General reasoning and generation | Open-ended text | $2/$10 per M tokens (reported) |
The Barbell Shape of the October Market
The wider AI model field looks like a barbell. On one end, labs keep pushing frontier scale, as with Mistral's Large 4 preview at roughly 1 trillion parameters or the four-tier rollout of GPT-6 inside ChatGPT. On the other, vendors ship cheap, narrow, fast models for the repetitive decision-making underneath those flagship assistants. Decision-1 sits firmly in the second camp, closer in spirit to a classifier than to a chatbot.
Which also means it is not really competing head-to-head with GPT-6 Sol or Gemini 4 Argon on the same job. It competes with them on cost and latency for the subset of requests that need no generation at all: the yes/no calls, the A/B/C classification, the pass/fail check. If a meaningful share of enterprise AI traffic is that mechanical, a cheap specialist model could pull a surprising amount of volume away from general-purpose APIs without ever writing a sentence of prose.
From Pre-Generative Classifiers to a Priced Product
Decision scoring is less an invention than a return to form. Before the generative AI boom, most production machine learning was exactly this: classifiers, rankers and scoring models trained for one narrow task. The generative wave of 2022 through 2025 pulled a huge share of engineering attention toward chatbots and agents able to do almost anything, at the cost of being slower and more expensive on simple work. Microsoft-Decision-1, AWS Strands Decider 2B and similar products mark an industry correction — use a frontier model when the task genuinely needs open-ended reasoning, and a small, cheap, fast model when it does not.
The LLM-as-judge pattern, where a language model evaluates or grades another model's output, has been a staple of AI evaluation pipelines for a couple of years. What changed in 2026 is packaging: vendors now sell that pattern as a dedicated, priced product rather than something developers hand-roll with prompt engineering on a general model. Microsoft's framing of Decision-1 for AI judging and agent guardrails places it squarely in that lineage, formalized and sold as infrastructure.
Who This Pressures
The most direct pressure lands on vendors selling general-purpose models for routing and classification tasks that customers are currently overpaying for relative to the job's actual complexity. If enterprises begin redirecting high-volume, low-complexity API calls toward purpose-built scoring models, that is a real dent in token volume for providers whose business model depends on charging full-model rates for work that does not need full-model reasoning.
It also puts competitive pressure on Google and Anthropic to decide whether they need their own named decision-scoring product, or whether they will fold similar functionality into existing lightweight tiers. Google has already been tightening its free-tier model lineup, visible in Gemini 4 Argon's pricing and benchmark split this year, while Anthropic's cost cuts to Claude Haiku suggest the same instinct: hold the frontier model's price up top and compete hard at the bottom of the market, where volume lives.
There is a cloud infrastructure angle too. Microsoft Foundry is Microsoft's platform for selling AI compute and model access to enterprise customers, and a cheap, high-volume, sticky workload like routing and classification is exactly the kind of traffic that keeps customers inside one ecosystem rather than shopping API calls across providers. A $0.042 per million input tokens product, if it performs as claimed, could become a quiet volume driver for Foundry even without becoming a consumer-facing name.
FAQ: What Readers Ask Next
- Is this a chatbot? No. Microsoft-Decision-1 returns a calibrated probability for each predefined option and generates no text, no explanations and no multimodal output; inputs are text up to 32,000 tokens.
- Have the speed claims been verified? Not independently in the coverage reviewed. The 35 times and 4.5 times multipliers, the approximately 35× P50 latency figure and Liberty Times' 85 milliseconds report trace back to Microsoft's benchmarking or launch materials.
- Why build on Alibaba's Qwen3.5-9B? Microsoft's stated plan is to rebase the model, including on Microsoft AI (MAI) and OpenAI, which reads as a speed-to-market choice on a narrow task rather than a permanent architecture decision.
- Can a team self-host it? No. Weights are not distributed, so access runs through hosted APIs on Microsoft Foundry or OpenRouter as microsoft/microsoft-decision-1, unlike the open Qwen3.5-9B base.
- What does it cost? The circulating figure from OpenRouter's listing is $0.042 per million input tokens with free output; Microsoft's own documentation reviewed here did not publish a confirmed price, and pricing is described as dynamic.
- Which checks should a pilot run first? Compare scores with past human decisions, apply a confidence floor such as 0.65, automate only above a chosen threshold such as 0.75, and treat a 90% prediction as expected to be right roughly nine times out of 10.
- Who else sells into this niche? Amazon's open-source Strands Decider 2B, reported at around 100 milliseconds in AWS testing, targets the same agent guardrails and routing space at a 2B parameter count.
What to Watch
Three follow-ups will decide whether this October launch becomes a durable product category or a footnote: whether anyone outside Microsoft reproduces the 36-benchmark comparison and the 35x-style latency claims; whether the promised rebasing from Qwen3.5-9B toward Microsoft AI (MAI) and OpenAI changes the accuracy and perturbation-stability numbers now advertised; and whether Google or Anthropic answer with their own named scoring tier. For buyers, the near-term question is narrower and more actionable — how much of today's AI spend is really fixed-option decision-making billed like generative reasoning.