Microsoft Decision-1 in Foundry: Fast AI Classification Model, Benchmarks and Pricing Caveats — Megathread

Microsoft launched a new AI model on Friday, and it is the strangest product the company has shipped in a year: a model that does not talk. Microsoft-Decision-1 does not generate text, hold conversations, or write code. You give it a fixed set of options, and it returns a calibrated probability for each one. That is the entire product. And it might be one of the most important AI releases this month, precisely because of what it is not.

What Microsoft-Decision-1 actually is​

The announcement came from Achint Srivastava, a Microsoft vice president, on the company's Command Line blog. Decision-1 is a 9-billion-parameter model post-trained from Qwen3.5-9B for what Microsoft calls single-pass decision scoring. It is available now in Microsoft Foundry and on OpenRouter, priced at $0.042 per million input tokens with output tokens free.
The interface is the point. Instead of a chat endpoint, Decision-1 exposes a structured Decisions API: you hand it a situation and a defined set of choices, and it scores each option. It handles yes-or-no decisions, multiple-choice selections, and rating scales, plus rubric-based grading of AI responses and agent actions. The probability is part of the API, not a side effect. A 90% score is supposed to mean the model is right about nine times in ten, and applications use that confidence to decide whether to act, defer, or send the case to a human for review.
Microsoft positions it for routing, classification, prioritization, verification, and workflow control. The worked examples from its own internal testing make the pitch concrete. The Xbox Research team used it to sort more than 10,000 open-ended game reviews and survey responses into researcher-defined themes, and found it competitive on quality with GPT-6 Sol while running more than 14 times faster at 200 times lower cost. The Copilot team uses it to grade chat and agentic responses, finding it competitive with GPT-5.6 Luna at 100 times the speed. Microsoft Discovery uses it for adaptive replanning in scientific experiments, where it scored as 46 times more consistent than an LLM-based judge at three times the speed.
The benchmark claims are aggressive. Microsoft says Decision-1 achieved the highest accuracy across 36 benchmarks spanning nearly 150,000 questions, kept blind from training, and was the fastest model measured: 2.5 times quicker than the runner-up H2O-Lightning-4B and 35 times quicker than GPT-6 Sol in P50 latency. On robustness, the company perturbs each request eight ways, through paraphrases, reordered options, and formatting noise, and reports the model flips its decision only 1.3% of the time, with zero flips when option descriptions are paraphrased or reversed.
Those are Microsoft's numbers, from Microsoft's tests, and they should be read that way until independent evaluations land. We will come back to that.

The real story: the unbundling of the LLM​

Step back from the spec sheet and the interesting thing is the category, not the model. For three years the industry's answer to every problem was a bigger general model: one system that writes, reasons, codes, and chats. Decision-1 is Microsoft betting the other way. Its own announcement frames it as choosing the right model for the right job, and the job here is narrow on purpose.
This is the unbundling of the large language model into a bench of specialists. The generators keep getting bigger and more expensive. But an agent that runs for hours, making hundreds of small operational calls, cannot afford to wake a frontier model every time it needs to decide whether to retry a step or hand off to a human. Each decision adds latency, and Microsoft's own example is telling: add 100 milliseconds to each of 20 sequential decisions and you have added two seconds to the workflow. At agent scale, the economics of every micro-decision start to dominate.
That is why the use case Microsoft keeps returning to is agent control. Evaluate an agent's proposed next step and decide whether to continue, stop, retry, or hand off to another model, tool, or human. Route an incoming request to the best model for the task based on quality, cost, and latency. Check a proposed interface action against safety policy before it executes. These are governor functions, and a governor needs to be fast, cheap, and calibrated, not eloquent. Nobody needs a sonnet about whether the support ticket goes to billing or engineering. They need a probability and a threshold.
Read that way, Decision-1 slots directly into the agentic platform story Microsoft has been building all fall. The October 7 keynote introduced Execution Containers to sandbox local agents, Hybrid Intelligence to move work between local models and the cloud, and a vision of agents acting on users' behalf with permission. Every one of those pieces needs a control layer: something that watches what agents propose and decides what happens next. A decision-scoring model is the natural shape of that layer. It is no accident that the use-case list in the announcement reads like a bill of materials for governing agents: AI judging, safety screening, skill-based decisions, computer and UI use.

Microsoft did not invent this category​

To Microsoft's credit, the announcement does not pretend otherwise. The appendix lists the models it benchmarked against, and the list is a tour of an emerging category: Quyet-1.0-Large, Surogate's Rune 26B, H2O's Lightning-4B, Strands' Decider 2B, Upstage's Solar Decide line, and notably GPT-6 Luna Decisions, which means OpenAI already ships a decisions API of its own with developer documentation. This is a recognized product category with multiple vendors, and Microsoft is entering it, not creating it.
That context matters for two reasons. First, it means the idea has survived contact with real customers at more than one company: there is genuine demand for a model that scores instead of chats. Second, it gives us something to watch. If decision models become the standard governor for agent workflows, the competition will be on calibration quality, latency SLAs, and price, not on who has the biggest parameter count. At $0.042 per million input tokens with free outputs, Microsoft is pricing like it wants volume.
There is also a personnel story worth one line. Srivastava co-founded Pi Labs, an AI evaluation startup, before joining Microsoft; Accel's profile lists Microsoft as the acquirer. A model for judging AI output, announced by the executive who used to sell AI evaluation tooling, is the kind of thing that makes sense in retrospect.

The skeptic's checklist​

Now the discipline. Everything impressive in the previous paragraphs comes from Microsoft's own materials, and a new model category deserves the same scrutiny we would give any vendor benchmark.
The benchmarks are self-reported. "Kept blind from training" is Microsoft's claim about its own process, and 36 benchmarks is a serious evaluation, but nobody outside the company has reproduced any of it yet. The competitive field is also small and young: being the most accurate decision model in late 2026 is a real result, but the category is months old, and the field will look different in a year.
Calibration is the load-bearing claim and the hardest to verify from a blog post. A probability API is only useful if the probabilities mean what they say on real production traffic, not just on benchmark sets. Microsoft reports strong calibration, but calibration has a way of degrading the moment the input distribution shifts, which in production it always does. The robustness numbers, 1.3% flip rate under perturbation, are genuinely impressive if they hold up, and also exactly the kind of number that needs independent replication.
The provenance is worth noting plainly. Decision-1 is post-trained from Qwen3.5-9B, an open model from a Chinese lab, and Microsoft says it will rebase future versions on other models including its own MAI models and OpenAI's. There is nothing wrong with that lineage, but it is worth knowing what is inside the box, and worth watching what changes when the rebase happens.
Finally, the safety evaluation, 5,250 requests across 11 benchmarks covering jailbreaks and prompt injection, is again Microsoft's own. A model whose job includes safety screening, deciding whether to allow, block, or escalate, needs adversarial testing by people who are not its vendor. That work has not happened yet in public.
None of this is a reason to dismiss the model. It is a reason to treat Friday's announcement as the opening of an evaluation period, not its conclusion.

Who should care​

If you are building agents, this is worth a look now, not later. The governor pattern, a cheap fast model watching a bigger model's outputs, is already the standard architecture for serious agent deployments, and most teams are currently implementing it with a general LLM they pay too much for and wait too long on. A purpose-built scorer at this price changes the math.
If you run IT for an organization deploying agents, the cost story is the headline. The Xbox example, competitive quality at 200 times lower cost than a frontier model for a classification workload, is the shape of the savings available across every high-volume decision point in an agent fleet: routing, triage, quality gates, escalation decisions. Those are exactly the workloads that quietly dominate AI spend once agents move past pilot stage.
If you are watching the platform story, Decision-1 is another piece of the scaffolding Microsoft is assembling around agents that act in the real world. Generators get the headlines. Governors get the control. The most important AI model this week might be the one that never writes a word.