Microsoft-Decision-1: 9B Qwen-based model scores answer options; 35x faster than GPT-6 Sol, Microsoft says

Microsoft-Decision-1 is a 9-billion-parameter model that does not write text. Given some content and a fixed list of answer options, it returns a calibrated probability for each option in one pass, so software can route a ticket, grade an agent's step or pick a model without parsing a chatbot's prose. Microsoft introduced it on 9 October 2026 in Microsoft Foundry and on OpenRouter, built it by post-training Alibaba's open-weight Qwen3.5-9B, and says it will "soon" rebase it on other models, including Microsoft AI (MAI) and OpenAI models. Every performance figure below is Microsoft's own; we found no independent reproduction.

Key points

• Microsoft-Decision-1 was announced on 9 October 2026 and is available in Microsoft Foundry and through OpenRouter as microsoft/microsoft-decision-1.
• It is post-trained from Qwen3.5-9B and returns a calibrated probability for each of a fixed set of answer options, not free-form text.
• Microsoft says it had the highest accuracy in a 36-benchmark comparison of nearly 150,000 held-out questions, and a P50 latency about 35 times lower than GPT-6 Sol and 2.5 times lower than the runner-up, H2O-Lightning-4B v1.1.
• OpenRouter lists a price of $0.042 per million input tokens and $0 for output, with about 33,000 tokens of context.
• The benchmark and cost claims are vendor-reported and have not been independently verified.

What happened

Microsoft's launch post, written by Achint Srivastava, VP of Software Engineering in the Office of the CTO, calls decision models "an important new category in AI": models "purpose-built to deliver structured outputs that software can immediately act on". Decision-1 supports yes/no, multiple-choice and rating options, and rubric-based grading of AI responses and agent actions, all through a structured API call. Microsoft names routing, classification, prioritisation, verification and workflow control as the target uses.

OpenRouter's model page, which links to Microsoft's announcement, adds that the model is "not intended for open-ended generation, conversation, translation, or summarization", that it runs on OpenRouter's Decisions API rather than the OpenAI-compatible chat endpoint, so chat-completions SDKs will not work with it, and that Azure is the only provider listed. It also says weights are "updated continually while the API shape stays the same". We found no sign that the weights themselves are downloadable; the page presents it as a hosted API only.

Background

Large language models have been used as classifiers and judges for years: you prompt a big model to answer "urgent or not?" and parse the text. That is slow, costs generation tokens, and gives you no trustworthy confidence number. Microsoft's pitch is that those calls are a large share of agent workloads and do not need a frontier model. Forrester analyst Charlie Dai told InfoWorld that routing, classification and approval checks often represent a large share of agent interactions, so a specialised model can cut cost and latency while leaving general-purpose models for reasoning and content generation. Microsoft's post makes the latency argument concrete: adding 100 milliseconds to each of 20 sequential decisions adds two seconds to a workflow.

Microsoft says it took several top models from the public JevBench leaderboard and tested them across 36 additional public and private benchmarks. An editor's note on the post says it was updated to add benchmarks for Jev on accuracy and calibration, which means the comparison set changed after launch. For readers following API costs, our explainers on how to estimate LLM API cost and how prices fell since June give the context for why a $0.042-per-million model matters.

How it works

Microsoft describes five design goals, each with a claim we can only report as the company's:

1. Speed. P50 latency about 35 times faster than GPT-6 Sol; in its benchmarking it was 2.5 times faster than the runner-up.
2. Quality that generalises. Accuracy measured on benchmarks kept blind from training, spanning routing, ranking, long context, multilingual and out-of-distribution tasks, reasoning and safety.
3. Robustness. Microsoft perturbs the same request eight ways (paraphrase, reordered options, formatting noise) and says the decision changed on 1.3% of perturbations on average, with zero flips when option descriptions were paraphrased or when options were reversed or shuffled.
4. Calibration. The probability is part of the API, so "a 90% prediction should be right about nine times out of 10 on representative cases". Applications use it to decide when to act, defer or ask for review.
5. Safety. Tested on 5,250 requests across 11 benchmarks covering harmful content, jailbreaks and prompt injection; Microsoft says it refused harmful behaviour while retaining a high degree of utility.

A practical pattern follows from calibration: act automatically above a threshold, send to a person or a bigger model below it. Microsoft's own use cases include agent controls (continue, stop, retry or hand off), model routing, AI judging, data labelling, code scanning and safety screening.

The numbers

Microsoft's internal tests, as described in the post:

• Xbox Research sorted more than 10,000 open-ended pieces of feedback into fixed themes. It found Decision-1 competitive with GPT-6 Sol on quality while running over 14 times faster and 200 times less expensive.
• The Copilot team, measuring chat and agent response quality, found it competitive with GPT5.6 Luna and 100 times faster.
• Microsoft Discovery, an adaptive-replanning agent, scored it 46 times more consistent than an LLM-based score at three times the speed, and nearly four times faster on replanning overall.

InfoWorld reports Microsoft's pricing example: an incident-prioritisation request of 86 input tokens and two output tokens costs about $0.0000036, so $1 would cover more than 275,000 such decisions. That arithmetic follows from the $0.042 per million input price.

What it means

For developers building agents, this is a cheap, fast "gate" component: classify, route and approve before calling a costly model. The trade-offs are real. Microsoft's own comparison pits a model that scores a fixed list against GPT-6 Sol, which generates text, so the 35x figure measures a P50 latency on Microsoft's benchmark, not a general speed advantage. Decision-1 gives a score but no rationale. Amit Kumar Jena of Kanerika told InfoWorld a precise-looking probability can give an uncertain decision more authority than it deserves and advised measuring "cost and latency per successfully completed workflow, including fallbacks and mistakes". Dai recommended calibration testing, controls on question wording and thresholds, audit logging and human review for high-impact decisions.

For the open-weight ecosystem, the story is also about provenance: a large US company has built a flagship launch on an open Chinese base model, and says it will later move to its own and OpenAI models. The hosted-only delivery means developers who assumed a Qwen derivative would be self-hostable should check before planning around it.

What to do if you want to try it

1. Start with reversible tasks such as triage and routing, as the analysts suggest.
2. Use the Decisions API shape (not chat completions) and define your answer options explicitly; wording of options affects results.
3. Build a labelled sample from your own data and check that "90%" really means about nine in ten before setting thresholds.
4. Log scores and keep a fallback to a larger model or a person below your threshold.
5. Re-test when Microsoft rebases the model, since behaviour may change.

What we don't know yet

• Independent results on accuracy, calibration and the 35x latency claim.
• Whether and when the model will be rebased, and how behaviour will change.
• Whether weights will ever be released.
• Pricing and regional availability beyond OpenRouter's listing (it is not available through EU in-region routing there).

What to watch next

Watch for third-party benchmark runs on JevBench and similar boards, Microsoft's rebased versions on MAI or OpenAI bases, and competing decision models from other vendors.

Sources

• Microsoft: Introducing Microsoft-Decision-1, our model for fast decision-making (primary, 9 October 2026)
• OpenRouter: Microsoft-Decision-1 model page
• InfoWorld: Microsoft separates decision-making from generative AI with a new model
• Shattered: Microsoft Decision-1 coverage