Microsoft Decision-1: New AI Model Scores Choices 35x Faster [2026]

Microsoft AI unveiled Microsoft-Decision-1 on October 9, 2026, a purpose-built “decision-scoring” model that the company says beats both general-purpose large language models and rival decision models on speed and accuracy. The release marks a notable shift in strategy for Microsoft: instead of chasing bigger, more expensive chatbots, it has built a small, fast model whose only job is to pick the right answer from a fixed list of options, and to do it in milliseconds rather than seconds.

The announcement, first reported by MarkTechPost and quickly picked up by outlets including Tech AI Magazine, TestingCatalog and AIBase, positions Microsoft-Decision-1 as infrastructure rather than a consumer-facing chatbot. It is meant to sit inside the plumbing of software systems, scoring choices for things like fraud flags, support-ticket routing, content moderation calls, and AI agent guardrails, rather than writing essays or holding conversations.

·

Google · Preferred Sources

Don't miss new tech stories on Google

Add TrendinTech once in the Google app and our stories appear in your news suggestions.

Add Now

What Microsoft-Decision-1 Actually Is

Microsoft-Decision-1 is built by post-training Qwen3.5-9B, the 9-billion-parameter open model from Alibaba’s Qwen family, according to Microsoft’s own announcement. That choice is itself notable: Microsoft chose an open-weight Chinese model as its starting point rather than building from scratch or starting with one of its own Microsoft AI (MAI) models. Microsoft has said it plans to rebase Microsoft-Decision-1 on other model families soon, including its in-house MAI models and OpenAI’s models, which suggests the Qwen3.5-9B base is a first iteration rather than a permanent architectural commitment.

The core design idea separates Microsoft-Decision-1 from almost every other model making headlines this year. Rather than generating free-form text token by token, the model takes a situation, a question, and a fixed set of answer options, then returns a calibrated probability score for each option in a single call. There is no essay, no chain-of-thought narration, no multi-turn back-and-forth. The system reads the inputs once and outputs a ranked set of scores that downstream software can act on immediately.

That narrow focus is the entire point. Microsoft is pitching Microsoft-Decision-1 for classification, routing, prioritization, verification, ranking, grading, workflow control, agent guardrails, and AI judging, tasks that currently eat enormous compute budgets when companies throw a full LLM at what is really a multiple-choice problem. If a customer support system just needs to know whether a ticket is “billing,” “technical,” or “cancellation,” running that query through a 400-billion-parameter chatbot is overkill. Microsoft-Decision-1 is designed to answer exactly that kind of question, fast and cheap.

Where You Can Get It and How It Ships

Microsoft-Decision-1 is available now through Microsoft Foundry, where the listing describes the model as being in public preview. That status matters for enterprise buyers: public preview typically means the model is usable in production-adjacent testing but Microsoft has not yet committed to the long-term service-level guarantees that come with general availability. Several outlets, including a writeup from OrcaRouter, have flagged that the Foundry listing shipped without a full published benchmark suite at launch, which has fed some early skepticism about how independently verifiable Microsoft’s performance claims are.

Microsoft also made Microsoft-Decision-1 available through OpenRouter, the multi-provider model marketplace, according to both Microsoft’s launch post and OpenRouter’s own listing. That is a meaningful distribution choice. OpenRouter is where a large share of independent developers and small AI startups go to compare models across providers without signing enterprise contracts, so putting Microsoft-Decision-1 there puts it in direct, easy reach of exactly the builders who are gluing together agent pipelines and need a cheap routing or verification layer.

On pricing, Microsoft’s official channels have not published a confirmed rate card as of this writing. A third-party source reported a rate of $0.042 per million input tokens with free output, a figure echoed in early community coverage such as a report from AlphaSignal, but that number has not been confirmed directly by Microsoft and should be treated as provisional until Microsoft publishes its own pricing page.

The Performance Claims, in Microsoft’s Own Words

Microsoft says Microsoft-Decision-1 topped a 36-benchmark comparison spanning nearly 150,000 questions, outperforming both general LLMs and other dedicated decision models on accuracy. On speed, the company reports the model ran 4.5 times faster than a competing system it calls Quyet-1.0-Large, and 35 times faster than GPT-6 Sol. Those are Microsoft’s own reported figures from its launch materials, not numbers that have been independently reproduced by a third-party benchmarking lab, and readers should weigh them accordingly, the same way any vendor-run benchmark deserves a grain of skepticism until outside researchers can test it themselves.

Microsoft CEO and Chairman Satya Nadella personally introduced the model in a public post, framing it as core enough to the company’s own operations to already be in internal use. “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,” Nadella wrote.

That quote is worth sitting with for a moment. Nadella did not call this a flagship release or a ChatGPT competitor. He called out three very specific, very operational uses: incident response, quality control, and scientific discovery. Those are not glamorous AI demo categories. They are the unglamorous, high-volume decision points that large organizations deal with thousands of times a day, and they are exactly the kind of workload that benefits most from a model that is cheap and fast rather than eloquent.

Why “Decision Models” Are Becoming Their Own Category

Microsoft-Decision-1 did not appear in a vacuum. It is part of a broader pattern across the AI industry in 2026: the splitting of “one giant model does everything” into a stack of specialized models, each tuned for a narrower job. The giant frontier chatbots, things like GPT-6 Sol and Claude Haiku 5.5, remain the tools of choice for open-ended reasoning, writing, and conversation. But running every single classification or routing decision through one of those models is slow and, at scale, expensive.

That gap is where decision-scoring models live. Rather than generating text, they are trained to output a probability distribution over a known, closed set of outcomes. This is conceptually closer to classic machine learning classifiers than to a chatbot, but Microsoft-Decision-1 keeps the flexibility of an LLM-trained backbone, meaning the same model can be reused across wildly different tasks just by changing the prompt and the option list, instead of training a brand-new classifier for every use case.

Quyet-1.0-Large, the model Microsoft says it beat by 4.5 times on latency, is itself evidence that other players are already building in this category. The existence of a named, benchmarked rival suggests decision-scoring is maturing into a recognized product category rather than a one-off Microsoft experiment. Enterprises running large-scale AI agent pipelines, think customer service bots that need to decide which of twelve possible actions to take next, are increasingly going to need a cheap, fast arbiter sitting between the user and the expensive reasoning model, and that arbiter is where Microsoft-Decision-1 is aimed.

Competitive Landscape: How Decision-1 Stacks Up

The table below lays out how Microsoft-Decision-1 compares with the systems Microsoft explicitly benchmarked against, using the figures from Microsoft’s own launch claims. These are vendor-reported numbers rather than independently audited results, but they are the only figures publicly available at launch.

Model	Type	Base Architecture	Relative Speed (P50 latency)	Primary Use Case
Microsoft-Decision-1	Decision-scoring model	Qwen3.5-9B (post-trained)	Baseline (fastest reported)	Classification, routing, grading, agent guardrails
Quyet-1.0-Large	Decision-scoring model	Undisclosed	4.5x slower than Decision-1	Structured decision tasks
GPT-6 Sol	General-purpose LLM	Undisclosed	35x slower than Decision-1	Open-ended reasoning and generation
Claude Haiku 5.5	General-purpose LLM	Undisclosed	Not benchmarked by Microsoft	Fast conversational and coding tasks

It’s worth noting that Claude Haiku 5.5 was not part of Microsoft’s published comparison, which only named Quyet-1.0-Large and GPT-6 Sol. It is included here purely for scale context, since the recent 90% Haiku price cut reshaped how teams think about “cheap but general” models, a different cost strategy from Microsoft’s “narrow but near-instant” approach.

Pricing Snapshot: What’s Confirmed vs. Reported

Detail	Status	Figure	Source
Input token pricing	Unconfirmed by Microsoft	$0.042 per million input tokens (reported)	Third-party report, cited by AlphaSignal
Output token pricing	Unconfirmed by Microsoft	Reported as free	Third-party report
Availability tier	Confirmed	Public preview on Microsoft Foundry	Microsoft Foundry listing
Secondary distribution	Confirmed	Listed on OpenRouter	Microsoft launch post; OpenRouter listing
Benchmark claim	Vendor-reported, not independently verified	Top accuracy across 36 benchmarks, ~150,000 questions	Microsoft announcement

Historical Context: Microsoft’s Long Bet on Small, Specialized Models

Microsoft-Decision-1 did not come out of nowhere strategically. The company has spent the last several years investing heavily in smaller, task-specific models under its Phi and MAI research lines, arguing that not every AI workload needs a frontier-scale model. That argument has only gotten stronger as inference costs have become a bigger line item for enterprises running AI at scale. A chatbot that costs a few cents per conversation is fine for a handful of daily queries. A routing decision made a million times a day by an automated system needs to cost fractions of a cent, or the economics simply do not work.

Microsoft’s relationship with OpenAI has also been evolving, and the plan to eventually rebase Microsoft-Decision-1 on OpenAI models, alongside Microsoft’s own MAI models, signals that the company wants optionality in its model supply chain rather than dependence on a single base architecture. Starting with an open-weight Qwen model from Alibaba, a company Microsoft does not have a deep commercial relationship with, is itself a sign of how commoditized base-model selection has become. The real competitive value, in Microsoft’s framing, is in the post-training and the decision-scoring architecture built on top, not in who trained the underlying weights.

This also fits into the broader pricing and efficiency battle playing out across the AI sector this year. Comparisons like GPT-6.1 Sol vs. Sonnet 5.5 vs. Gemini 3.8 Flash have shown roughly a 3x price gap between frontier model tiers, and that same cost pressure is what makes a dramatically cheaper, narrower model like Microsoft-Decision-1 attractive to engineering teams trying to control their AI spend without giving up accuracy on the decisions that matter.

Market Impact: Who Should Be Paying Attention

The most immediate audience for Microsoft-Decision-1 is not consumers, it is platform engineers and AI infrastructure teams building agentic systems. Anyone running an AI agent pipeline today already knows the pain point Microsoft is targeting: every agent needs some mechanism to decide what to do next, whether to escalate, whether a generated answer passed a safety check, or which of several tools to call. Today, many teams solve this by calling a full LLM and asking it to pick from a list, which is slow and burns tokens on a task that does not need free-form generation.

If Microsoft’s reported 35x speed advantage over GPT-6 Sol holds up under independent testing, it has real implications for latency-sensitive applications. Developer tools and coding assistants are a good example. Platforms that already combine multiple AI models in a single workflow, the kind of pattern covered in the Cursor and Claude Code integration workflow, could plausibly swap in a decision-scoring layer for intermediate steps like deciding whether a code suggestion needs a second pass, without touching the main reasoning model.

There is also a cloud infrastructure angle. Companies already weighing serverless and inference costs across providers, the kind of comparison laid out in AWS Lambda vs. Azure Functions vs. Cloud Run pricing, now have a new variable to factor in if they run workloads on Azure: a Microsoft-native decision layer that, if the reported pricing holds, could meaningfully undercut the cost of running classification tasks through a general-purpose model hosted on the same cloud.

For competitors, Microsoft-Decision-1 raises a question that OpenAI, Google, and Anthropic will all need to answer: do they build their own narrow decision-scoring products, or do they bet that frontier models with falling prices, like the trend seen in the recent AI chatbot benchmark comparisons, will make specialized decision models unnecessary. Microsoft is betting that specialization wins on cost and latency even as general models get cheaper, because the gap between “cheap general model” and “purpose-built narrow model” doesn’t close, it just moves.

What Microsoft Hasn’t Confirmed Yet

Several details remain open as of this writing. Microsoft has not publicly named a specific executive beyond Nadella’s launch post associated with the Microsoft-Decision-1 release, so readers should be cautious about any secondary reporting that attributes detailed technical commentary to unnamed “Microsoft officials.” Whether Microsoft-Decision-1’s model weights will be released publicly, in the way some of Microsoft’s Phi models have been, has also not been established consistently across early coverage, and Microsoft’s own Foundry listing does not clarify the point. Final, Microsoft-confirmed pricing has not been published; the $0.042-per-million-token figure circulating in early coverage originates from third-party reporting, not from an official Microsoft pricing page, and could change before or after the model exits public preview.

These gaps matter because they affect whether Microsoft-Decision-1 is ready for serious production commitments today or is still best treated as an evaluation-stage tool. Public preview status on Foundry typically signals that Microsoft itself views the release as not yet fully production-hardened, and the absence of a complete, independently reproducible benchmark suite at launch, flagged by outlets including OrcaRouter, means outside developers are currently relying on Microsoft’s own numbers rather than a neutral third-party evaluation.

Predictions: Where Decision-Scoring Models Go From Here

• Expect Microsoft to publish a formal, Microsoft-confirmed pricing page for Microsoft-Decision-1 within weeks of this public preview launch, likely aligning close to the $0.042-per-million-token figure already circulating, since Microsoft rarely lets unconfirmed third-party pricing stand uncorrected for long.
• Expect at least one of OpenAI, Google, or Anthropic to announce a comparable narrow decision-scoring or classification-focused model within the next two quarters, following the same logic that drove Microsoft’s move: agent pipelines need a cheap arbitration layer, and none of the frontier labs want to cede that layer to a rival’s cloud ecosystem.
• Expect the planned rebasing of Microsoft-Decision-1 onto Microsoft’s own MAI models and OpenAI models to arrive before Microsoft-Decision-1 exits public preview, since Microsoft explicitly flagged this as a near-term roadmap item rather than a distant goal.
• Expect independent benchmarking groups to publish their own comparisons of Microsoft-Decision-1 against Quyet-1.0-Large and general LLMs within the next few months, which will either validate or meaningfully soften Microsoft’s reported 35x speed advantage over GPT-6 Sol.
• Expect enterprise AI agent platforms, including the growing ecosystem of coding and automation tools, to begin advertising “decision-scoring layer” integrations as a selling point by early 2027, mirroring how retrieval-augmented generation and function-calling became standard marketing checkboxes in prior product cycles.

Why This Matters Beyond the Headline

It is tempting to read Microsoft-Decision-1 as just another incremental model release in a year that has already produced an overwhelming number of them. But the strategic signal is bigger than the model itself. Microsoft is explicitly telling the market that not every AI problem is a chatbot problem, and that there is real commercial value in building infrastructure-grade models that nobody will ever talk to directly. If that bet pays off, Microsoft-Decision-1 could end up quietly running behind the scenes of far more software systems than any flashy new chatbot release, simply because routing, grading, and verification decisions happen far more often than open-ended conversations do.

The unanswered questions, pricing confirmation, weight availability, and independent benchmark validation, are exactly the things worth watching over the next few months. Until Microsoft locks those down, Microsoft-Decision-1 remains a genuinely interesting bet rather than a settled product, and the gap between “Microsoft says” and “independently confirmed” is where the real story will play out.

Frequently Asked Questions

What is Microsoft-Decision-1?
Microsoft-Decision-1 is a decision-scoring AI model released by Microsoft on October 9, 2026. Instead of generating free-form text, it scores a fixed set of predefined answer options and returns a calibrated probability for each one, making it suited to classification, routing, grading, and AI agent guardrail tasks.

What model is Microsoft-Decision-1 based on?
Microsoft says it post-trained Microsoft-Decision-1 from Qwen3.5-9B, the 9-billion-parameter open model from Alibaba’s Qwen family. Microsoft has also said it plans to rebase the model on its own MAI models and on OpenAI models in the future.

Where can I access Microsoft-Decision-1?
The model is available now in public preview on Microsoft Foundry, and it is also listed on OpenRouter, according to Microsoft’s launch post and the OpenRouter listing itself.

How much does Microsoft-Decision-1 cost?
Microsoft has not published official confirmed pricing. A third-party source reports a rate of $0.042 per million input tokens with free output, but this figure has not been confirmed directly by Microsoft and should be treated as provisional.

How fast is Microsoft-Decision-1 compared to other models?
Microsoft reports that Microsoft-Decision-1 ran 4.5 times faster than Quyet-1.0-Large and 35 times faster than GPT-6 Sol at typical latency. These are Microsoft’s own reported benchmark figures and have not yet been independently reproduced by outside researchers.

Is Microsoft-Decision-1 meant to replace chatbots like ChatGPT or Claude?
No. Microsoft positions Microsoft-Decision-1 as infrastructure for narrow, structured decisions, not as a general conversational assistant. It is meant to work alongside general-purpose LLMs, handling routing and classification tasks that do not require open-ended text generation.

Who is already using Microsoft-Decision-1?
Satya Nadella said Microsoft is already testing the model internally across use cases including incident response, quality control, and scientific discovery. Microsoft has not named external customers using the model as of this announcement.

Is Microsoft-Decision-1 generally available or still in preview?
As of the October 9, 2026 announcement, the Microsoft Foundry listing describes Microsoft-Decision-1 as being in public preview, not full general availability.

Related Coverage

• Snapdragon X2 vs Core Ultra 3: 24% Faster, $600 Less [2026]
• RTX 5080 vs RX 9070 XT: 38% Faster, $400 Pricier [2026]
• Best AI Image Generators: $4.99 to $120 [2026]
• Gigabyte BIOS Hints at Intel 2027 CPUs With Dual RAM [2026]
• Best Mechanical Keyboards 2026: 13 Steps, 45 Min