Microsoft launched Microsoft-Decision-1 in Microsoft Foundry on 9 October 2026, with availability on OpenRouter following. The model is, in Microsoft’s own description, “purpose-built to deliver structured outputs that software can immediately act on.” Rather than generating text, it takes a fixed set of answer options — yes/no, multiple choice, or a rating scale — and returns a calibrated probability score for each one through a structured API call. Microsoft lists its intended uses as routing requests, classifying inputs, verifying outputs, prioritising work, and rubric-based grading of other AI models’ responses and agent actions.
The model was post-trained from Alibaba’s Qwen3.5-9B, and Microsoft says it plans to rebase future versions on its own MAI models and on OpenAI technology instead. Microsoft prices input tokens at 0.042 dollars per million, and makes output tokens free, reflecting that the model’s output is a short structured score rather than generated prose.
Why it matters
Most production AI systems now have a layer of smaller models doing routing, classification, and grading work behind the scenes — deciding which larger model should handle a request, checking an agent’s output against a rubric, or verifying that a tool call was safe to execute. That layer is usually built by running a general-purpose LLM on a narrow task, which is both slower and more expensive than the task requires. A model purpose-built to return a calibrated probability over fixed options, rather than generate and then parse text, targets that specific cost and latency bottleneck directly.
Microsoft is not alone in building this: TypeSafe, maker of a competing decision-scoring model called Jev, raised 870 million dollars in a Series A round on 10 October 2026 at a 7.5-billion-dollar valuation, suggesting investors already see this as a distinct product category rather than a Microsoft-only idea.
What Microsoft’s benchmark numbers actually measure
Microsoft says Microsoft-Decision-1 reached the highest accuracy in a comparison across 36 benchmarks spanning nearly 150,000 questions, which it says were withheld from the model’s training data. The benchmarks cover routing, ranking, long-context decisions, multilingual classification, reasoning, out-of-distribution tasks, and safety evaluation. On this internal test suite, Microsoft reports Decision-1 scored an average accuracy of 83.5 percent against 82.3 percent for the next-best system, a decision model called Jev 1.13.0. Microsoft also reports a median response latency of 85 milliseconds, with a 95th-percentile latency of 125 milliseconds.
The latency comparisons Microsoft is most publicly promoting are less apples-to-apples than they first appear. Microsoft says Decision-1 is 2.5 times faster than the runner-up decision model, H2O-Lightning-4B v1.1, at the 50th-percentile latency — a reasonable like-for-like comparison, since both are purpose-built scorers. But the headline figure in Microsoft’s own marketing, that it is 35 times faster than GPT-6 Sol, compares a narrow, single-purpose scoring model against a full general-purpose generative model that was never built to do only this one task quickly. That comparison is true as stated but measures two different kinds of system, not two competitors in the same category.
On robustness, Microsoft reports that across eight paraphrased or reordered variants of the same question, Decision-1 changed its answer 1.3 percent of the time on average, with zero changes when only the wording of an option’s description was paraphrased. Microsoft also says it ran 5,250 adversarial requests across 11 safety benchmarks testing for harmful content and jailbreak susceptibility, though it has not published the raw pass/fail rate from that safety evaluation, only that it conducted it. Every number in this section is Microsoft’s own, run on its own choice of benchmarks and comparison models; none of it has been independently reproduced.
What this means for buyers
Teams currently spending a general-purpose LLM’s token budget on routing, classification, or grading subtasks have a concrete reason to benchmark a purpose-built scorer like Decision-1 against their own workload, given the free output-token pricing and the plausible latency gains for genuinely fixed-choice decisions. The comparison that matters for a buyer’s own evaluation is against other decision-scoring models — Jev, H2O-Lightning, and similar systems — not against the 35-times-faster-than-GPT-6-Sol framing Microsoft leads with, which compares against a tool not built for this narrow job.
Buyers should also plan around the base-model change Microsoft has already announced: because Decision-1 is currently built on Qwen3.5-9B but will reportedly move to Microsoft’s own MAI and OpenAI-derived models, scores and latency on today’s version may not hold once that migration happens, following the same rollout pattern 18BYTE tracked with OpenAI’s GPT-6 rollout, where the underlying model changed under an interface users had already adopted.
What would change our reading
This would move toward a stronger endorsement if an independent third party reproduced Microsoft’s 36-benchmark accuracy comparison, or if Microsoft published the raw pass/fail numbers from its 5,250-request safety evaluation rather than only the fact that the evaluation happened. It would move the other way if early adopters reported accuracy or latency on real workloads that diverged significantly from Microsoft’s internal figures, which is common with narrow benchmark suites that do not match production traffic patterns. We will also watch what happens to existing Decision-1 deployments once Microsoft rebases the model away from Qwen3.5-9B onto its own models, since that is a confirmed future change to a system buyers are being asked to adopt today.