Microsoft-Decision-1.1, Microsoft's Decision Model, Arrives, Outperforming GPT-6 Luna and Gemini
IT Home, November 18: Microsoft announced yesterday (November 17) the official launch of the Microsoft-Decision-1.1 decision AI model, which is now available on Microsoft Foundry, OpenRouter, and Vercel.

The model targets high-frequency decision scenarios in enterprise software, helping applications quickly make structured judgments in tasks such as routing, classification, prioritization, content verification, and workflow control.
In terms of pricing, Microsoft Foundry's Global Standard input tokens are priced at $0.042 per million tokens (IT Home note: approximately 0.28 yuan at current exchange rates), with output tokens free; the U.S. data zone, EU, and Asia-Pacific data zones are priced at $0.046 (approximately 0.31 yuan) and $0.05 (approximately 0.33 yuan) per million tokens, respectively.
Unlike traditional general-purpose large language models, Microsoft-Decision-1.1 focuses on choosing from allowed actions within rules and constraints set by developers. It also outputs calibrated confidence scores; when information is insufficient or uncertainty is high, it can choose to "abstain" or prompt for human intervention, thereby reducing the risk of AI errors in critical business processes.

Microsoft says the new version improves both intelligence and calibration, and while maintaining low cost, it is more than twice as fast as the previous generation Microsoft-Decision-1.
On the Artificial Analysis Decision Index v2.1 benchmark, the model scores 68, 6 points higher than the previous generation, and leads models including Jev 1.14, Google Gemini Decision 1, and OpenAI GPT-6 Luna (Decisions).
Microsoft-Decision-1.1 ranks eighth in the decision model rankings. In the Artificial Analysis Decision Index, Microsoft-Decision-1.1 is 2 points higher than Elephant 9B and Jev 1.14, but 1 point lower than d1.2. IT Home attaches the relevant screenshots below:


To balance real-time performance and cost, Microsoft built a custom inference engine for the model and optimized it based on the mid-sized mixture-of-experts model Gemma 4 26B A4B. The model can run on various GPU hardware such as A100, H100/200, and GB200/300, and reduces inference costs by intelligently scheduling requests.

Microsoft also said the model's training combines supervised fine-tuning and reinforcement learning, using synthetic and licensed datasets covering tens of thousands of task types to improve its generalization in real business scenarios.
The model now supports multimodal input and can receive image information such as photos and screenshots. Since its public preview release, more than 4,000 customers have used it, with application scenarios including model routing, customer service ticket triage, business process automation, content tagging and classification, among others; early customers include Epic, Equinor, Atos Group, and Raiffeisen Bank.
Within Microsoft, the model has been used for Copilot request routing: its judgment accuracy is comparable to large language models, but it is 2.2 times faster and 4.4 times cheaper. It is also used for dynamically adjusting task reasoning intensity in Cowork, and for quality and safety evaluations in Foundry Evaluations.
Related articles
- Microsoft and Apple outages in succession; Google responds urgently to Gmail outage rumors, saying "services are operating normally"
- Multilingual error rate drops by 50%: Microsoft upgrades doctor-patient conversation AI transcription engine, now covering 1/3 of physicians in the U.S.
- Microsoft's first speech-to-speech model: MAI Realtime 1 arrives, starting at $0.01 per minute
- Microsoft's most cost-effective cybersecurity AI model, MAI Cyber 2 Flash, arrives: performance comparable to GPT-6 Astra, at about 2% of its cost per task
- Microsoft announces PowerShell 7 will ship with Windows by default; Win11 26H2 will be the last version with 5.1 pre-enabled
- "World's No. 1 image editing AI model": Microsoft MAI Image 2.7 arrives, tops GPT Image 2.6 at 1/5 the price