Microsoft Releases Microsoft-Decision-1.1, More Than Doubling Speed and Adding Image Input
Microsoft has released its next-generation decision model, Microsoft-Decision-1.1, which runs more than twice as fast as its predecessor, supports image input, and scored 68 on the Artificial Analysis Decision Index. Microsoft says the model achieved the highest average accuracy among eight decision models across 102 benchmarks. When used to route Copilot requests, it delivered accuracy comparable to that of large language models while running 2.2 times faster and costing about one-quarter as much.
Why read this
AI score 72
The materials also provide third-party rankings, Microsoft’s internal test results, and product-routing data, enabling comparisons of efficiency differences between specialized decision models and general-purpose large language models.
Topics
Body · AI translation
November 18 news from Zhidx: Microsoft today released its next-generation decision model, Microsoft-Decision-1.1. It scored 68 on the Artificial Analysis Decision Index, ranking eighth—six points higher than Microsoft’s first decision model and ahead of Jev 1.14, Google’s Gemini Decision 1, and OpenAI’s GPT-6 Luna decision model.
On October 10, Microsoft launched its first proprietary decision model, Microsoft-Decision-1. Microsoft says that, compared with its predecessor, Microsoft-Decision-1.1 offers greater intelligence, more precise calibration, more than twice the processing speed, and lower costs.
In addition to the decision model, Microsoft unveiled at its Ignite conference the MAI Code 2 and MAI Code 2 Flash coding models, designed for routine enterprise coding tasks, as well as the new Code 2 and Code 2 Flash models, built specifically for programming and cybersecurity. Across three coding benchmarks measuring routine, high-volume coding tasks, MAI Code 2 Flash delivered performance comparable to GPT-6 Luna and Claude Haiku 5.5 at just 65% and 20% of their respective costs.
The next-generation decision model adds multimodal support and can accept images. In Microsoft Foundry, it is priced at $0.042 per million input tokens, with output free of charge.
Microsoft says its decision models are designed to make calibrated decisions under uncertainty across a wide range of real-world software use cases. Within constraints set by developers, the models select from permitted actions and output calibrated confidence scores, allowing applications to defer decisions when uncertainty is high.
Moreover, because these models focus specifically on decision-making, they are significantly faster and less expensive than general-purpose large language models when handling decision-related tasks.
Microsoft pursued joint innovation across latency, cost, intelligence, and calibration:
Latency: Decision models must operate in real time, even when handling complex requests with long inputs and hundreds of questions. Unlike conventional large language model inference engines optimized for token-by-token text generation, Microsoft-Decision-1.1 uses a custom inference engine optimized for the computational requirements of decision models.
Cost: Microsoft’s custom inference engine is optimized for low cost by efficiently distributing requests across GPU clusters, rather than solely pursuing low latency. The model was developed on the medium-sized mixture-of-experts model Gemma 4 26B A4B. Gemma 4 26B A4B has enough active parameters to provide the intelligence required for decision-making while remaining efficient enough to run on hardware in a variety of flexible configurations, including A100, H100/200, and GB200/300 systems.
Intelligence: Decision models must remain small enough to operate across all scenarios, generalize well to entirely new or unseen tasks, and be trained specifically to make decisions. To achieve this, Microsoft combined supervised fine-tuning with reinforcement learning, using hundreds of synthetic and licensed datasets covering tens of thousands of task types across numerous industries. This trains the model to exercise judgment rather than memorize answers, enabling it to distinguish between highly similar options, recognize uncertainty, and determine clearly when to answer and when to decline.
Calibration: Microsoft trained the model to understand the boundaries of its own knowledge, provide consistent answers in similar scenarios, accurately report confidence in its responses, divide confidence among multiple valid answers, and decline to answer when information is incomplete. Developers can set their own thresholds for making decisions, flagging items, or escalating them for further handling.
In addition to third-party testing, Microsoft evaluated Microsoft-Decision-1.1 on 102 benchmarks covering more than 385,000 questions it had not encountered during training. The test tasks spanned routing, ranking, long-context processing, multilingual capabilities, out-of-distribution tasks, reasoning, safety, and other areas.
Microsoft-Decision-1.1 achieved the highest average accuracy among eight decision models and led in seven of nine capability categories.
In just over five weeks since launch, more than 4,000 customers have used Microsoft’s decision models for tasks including model routing, customer-support triage, business-process automation, labeling, and classification. Early adopters include Epic, Equinor, Atos Group, and Austria’s Raiffeisen Bank.
At the same time, Microsoft is integrating Microsoft-Decision-1.1’s specialized intelligence into a range of products to enable fast and efficient decisions, while reserving larger models for more complex tasks.
In Copilot, Microsoft-Decision-1.1 routes user requests to the appropriate interaction environment, determining whether Copilot or Cowork is best suited to handle them. Its routing accuracy is comparable to that of a large language model, while it is 2.2 times faster and costs about one-quarter as much.
In Cowork, the model powers Auto, dynamically adjusting the amount of reasoning effort Cowork devotes to each task.
Conclusion: Microsoft’s models are deeply integrated with its own products, potentially charting a differentiated path
Unlike other leading model developers, Microsoft distinguishes itself through the deep integration of its models with its own products, including GitHub Copilot, Foundry, and Copilot workflows. Model capabilities are embedded directly into existing productivity and development interfaces rather than being sold solely through standalone APIs.
This means enterprises can directly integrate a complete solution into their existing processes, with clearly defined compliance boundaries and lower adoption costs.
AI Trends