Published: 12 October 2026
Microsoft released Microsoft-Decision-1 on 9 October 2026. The model does not write text. It scores a fixed set of answer options and returns a probability for each one. It costs $0.042 per million input tokens with free output, the same price as TypeSafe’s Jev, the model every new decision model is chasing. Microsoft calls it the most accurate model it tested, but its announcement shows no scores.
Microsoft claims the accuracy lead and only shows speed
A decision model answers closed questions inside an AI agent: where to route a ticket, whether to allow an action, how to rate an answer. Microsoft built Decision-1 on Alibaba’s open Qwen3.5-9B and says it will later move it onto its own MAI models and OpenAI technology. It runs in Microsoft Foundry and through OpenRouter.
Nearly everything Microsoft put in numbers concerns price and speed, not accuracy.
| What is measured | Microsoft-Decision-1 | Alternative |
|---|---|---|
| Input price (per million tokens) | $0.042, output free | Jev: $0.042; Clef-flash: $0.038 |
| Context window | 32,768 tokens | Clef: 64,000 tokens hosted |
| Input types | Text only | Clef-omni: text, images, audio, video |
| Median speed (P50) | About 35x faster than GPT-6 Sol | 2.5x faster than runner-up H2O-Lightning-4B v1.1 |
| Stability (8 rewordings of one request) | Answer changed 1.3% of the time | No rival figure published |
| Accuracy over 36 benchmarks | “Highest”, no score given | No rival scores given |
All Decision-1 figures are Microsoft’s own. The post does not describe the hardware or setup behind the speed claims. An editor’s note says Jev results were added later, but no Jev numbers appear in the text. The rival prices come from Cloudflare’s 9 October announcement, not from Microsoft.
Microsoft’s model card is unusually clear on one point: Decision-1 should not be the only decision-maker in credit, hiring, housing, insurance, education, healthcare or legal matters. It also names no Foundry regions, so it is unclear where your data is processed.
If your agents already run on Azure, this is the easiest decision model to try, because it sits next to the models you use today. Pick one routing or guardrail step where you know the right answers, and check whether its probabilities match what really happened. Until Microsoft publishes accuracy and calibration numbers, treat the accuracy claim as untested. Teams outside Azure have little reason to switch: Clef-flash is slightly cheaper, Strands Decider 2B runs on your own hardware, and both publish their scores.
For more information, visit the official announcement of Microsoft-Decision-1 on the Microsoft blog.
