# Microsoft-Decision-1 Microsoft-Decision-1 is a [[Decision Models (DMs)|decision model]] that [[Microsoft]] announced on 9 October 2026. You can deploy it from the Microsoft Foundry model catalog or call it through [[OpenRouter]]. Like [[Jev]] and [[Cloudflare Clef]], it never writes text. You send it some state (text or JSON) plus typed questions, and it returns a probability for every allowed answer. Microsoft pitches it for routing, classification, prioritization, verification and workflow control, both in regular applications and inside [[AI Agents]]. Why does it matter? Jev launched on 15 September. Less than four weeks later, Microsoft shipped its own decision model with the same API shape and exactly the same price. OpenAI is in the game too (its "GPT-6 Luna Decisions" shows up in Microsoft's comparison, see [[GPT-6 Luna]]). So decision models now have a hyperscaler behind them, and they come with the things enterprise teams already rely on in Foundry: Entra ID, quotas, data zones and Azure billing. ## How it works - **Base model.** Microsoft post-trained [[Qwen]]3.5-9B for fast, single-pass decision scoring (the same backbone Cloudflare used for Clef-flash). Microsoft says it'll soon rebase the model on others, including Microsoft AI (MAI) and [[OpenAI]] models - **Question types.** The [[System One Primitives]] under the same names: `noul` (a yes/no probability between 0 and 1), `choice` (one option from a fixed set, with a probability per option) and `score` (a level on an ordered scale). The score is the probability-weighted average of the level indexes, so on a four-level scale (0 to 3) the result can land between two levels. Microsoft also lists rubric-based grading of AI responses and agent actions - **API.** Requests go to `<your-foundry-endpoint>/providers/microsoft/v1/systemone`, and the docs describe the request pattern as "similar to a system-one API". Several questions about the same state can share one request. Auth is Entra ID (recommended) or an API key - **Deployment.** Model version 1, as GlobalStandard or as DataZoneStandard in selected regions (inference stays inside the chosen data zone). On OpenRouter, it's `microsoft/microsoft-decision-1` - **Price.** $0.042 per million input tokens; output tokens are free. That's Jev's price to the cent ## Numbers All of these come from Microsoft, so read them as vendor benchmarks. - **Accuracy:** 83.5% average over 36 benchmarks (147,137 questions, kept blind from training), first place. Next come Jev (82.3), Quyet-1.0-Large (81.9, #1 on the JevBench leaderboard on 8 October), Surogate Rune 26B-A4B (79.7), GPT-6 Luna Decisions (79.4), deck-31B (77.8) and H2O-Lightning-4B (77.2). AWS's Strands-Decider 2B scored 54.8, but only on the 23 benchmarks it could answer - **Latency:** 85 ms at the median, 125 ms at p95. That's 2.5x quicker than the runner-up (H2O-Lightning-4B, 210 ms), against 240 ms for Jev and 3.01 s for [[GPT-6 Sol]] (~35x) - **Calibration:** 92.2, third behind Jev (93.7) and Quyet (93.1). Microsoft insists that the probability is part of the API: a 90% prediction should be right about nine times out of ten ([[AI Model Calibration]]) - **Rewording and reordering:** Microsoft perturbs each request in eight ways (paraphrased instructions, reordered choices, renamed keys, formatting noise...). The decision flips on 1.3% of perturbations on average, and never when option descriptions are paraphrased or options are reversed or shuffled - **Safety:** tested on 5,250 requests across 11 benchmarks (harmful content, jailbreaks, prompt injection). Microsoft says it refused harmful behavior "while retaining a high degree of utility", without publishing the figures - **Cost:** classifying game reviews and livestream posts took 143 to 188 ms per text, against 2.6 to 2.8 s for GPT-6 Sol, at roughly 1/200th of the cost. Scaled to a million texts: about $11 instead of about $2,434 ## How Microsoft uses it internally - **Xbox Research** sorted more than 10,000 pieces of player feedback (surveys, Steam, X) into a fixed set of themes. Microsoft says the quality was competitive with GPT-6 Sol, at 14x the speed and 1/200th of the cost - **The Copilot team** scores the quality of chat and agentic responses with it: competitive with GPT-5.6 Luna and 100x faster - **On-call engineers** use it for knowledge retrieval during live incidents; Microsoft says it beat an LLM on quality and speed, with no numbers - **Microsoft Discovery** uses it to grade experiments with a rubric during adaptive replanning: 46x more consistent than the LLM-based score, at three times the speed, and almost four times faster replanning overall The list of suggested uses is long: [[Model routing]], agent controls (continue, stop, retry, or hand off to a model, a tool or a human), AI judging, intent analysis, data validation, search relevance, code scanning, safety screening ([[AI Guardrails]]), computer use and even robotics. The one I find the most interesting is "skill-based decisions": apply the rules of an agent skill to pick the next action without making the main model reread a long list of instructions every time. ## Watch out for - **Microsoft picked the benchmarks.** The 36 benchmarks mix public and private sets. There's no paper, no model card with training details, and no weights. It's API-only, and Microsoft doesn't say how the post-training was done - **The latency comparison isn't like for like.** Microsoft-Decision-1 was measured through Foundry, in the same region as the client. The other models' figures are JevBench's adjusted medians - **Jev was added later.** An editor's note says the post was updated to add Jev's accuracy and calibration. So the first version compared Microsoft-Decision-1 with everything except the model that defined the category. With Jev included, the accuracy lead is 1.2 points, and Jev is better calibrated - **Beating GPT-6 Sol on speed proves little.** A 9B single-pass model will always answer faster than a frontier LLM that generates text. The comparison that matters is with other decision models, and the gaps there are small - **Microsoft's own docs list the limits.** Scores can change with how you phrase or order the questions and options (the perturbation test above only counts decision flips, so a score can move while the pick stays the same). Calibration is strongest on familiar task types. The model may reflect biases from its base model and training data, gives no explanations, may rely on outdated knowledge, and as a safety filter can miss subtle harmful content or flag benign content. Their advice: validate on your own data, set thresholds based on the cost of false positives and false negatives, add an abstention option, randomize option order, and keep meaningful human review for consequential decisions about people (credit, employment, housing, healthcare, legal) ## My take For me, the API is the most telling part of this launch. The endpoint is literally called `systemone`, the question types keep TypeSafe's names (`noul`, `choice`, `score`), and the price matches Jev's exactly. Microsoft didn't design a competing interface. It adopted the one Jev introduced. After Clef, that makes "state in, typed probabilities out" a de facto standard, and code written for one provider should move to another without much work. So the competition moves elsewhere: calibration, training data, and who already sits in the request path. Cloudflare has the edge network. Microsoft has every company that already runs on Azure, and for them a decision model is now one more deployment in a project they already have. Would I switch from Jev? Not on these numbers alone. The accuracy lead is small, Jev calibrates slightly better, and calibration is what you rely on when you [[Confidence-Gated Routing|gate on a threshold]]. I'd run Microsoft-Decision-1, Jev and Clef-flash side by side on my own labeled examples ([[Shadow Evaluation]]) and pick per task. And it's good news either way: a third serious vendor at $0.042 per million tokens means nobody's decision layer has to depend on a single closed API anymore. ## References - [Introducing Microsoft-Decision-1, our model for fast decision-making (Microsoft Command Line, 2026-10-09)](https://commandline.microsoft.com/microsoft-decision-1-model-foundry/) - [Deploy and use Microsoft-Decision-1 in Microsoft Foundry (Microsoft Learn)](https://learn.microsoft.com/en-us/azure/foundry/foundry-models/how-to/use-foundry-models-microsoft-decision) - [Microsoft-Decision-1 in the Microsoft Foundry model catalog](https://ai.azure.com/catalog/models/Microsoft-Decision-1) - [Microsoft-Decision-1 on OpenRouter](https://openrouter.ai/microsoft/microsoft-decision-1) - [JevBench leaderboard (as linked by Microsoft)](https://benchmarkheaven.com/jev-models) ## Related - [[Decision Models (DMs)]] - [[Decision Models (MoC)]] - [[Jev]] - [[Cloudflare Clef]] - [[TypeSafe AI]] - [[System One Models]] - [[System One Primitives]] - [[Microsoft]] - [[Qwen]] - [[OpenRouter]] - [[GPT-6 Sol]] - [[GPT-6 Luna]] - [[Model routing]] - [[Confidence-Gated Routing]] - [[AI Model Calibration]] - [[AI Guardrails]] - [[Shadow Evaluation]] - [[AI Agents]] - [[Large Language Models (LLMs)]]