Back to Blog

Microsoft Decision 1.1: 2x Faster, $0.042/M Input

Daily News6398
Microsoft Decision 1.1: 2x Faster, $0.042/M Input

Microsoft rolled out Microsoft-Decision-1.1 on November 17, bringing a purpose-built AI model optimized for structured decision-making to developers. The model is now live on Microsoft Foundry, OpenRouter and Vercel. The company positions it as a low-cost, high-speed solution, with stronger reasoning calibration and more than double the inference speed of its prior generation.

Unlike general-purpose large language models that focus on open-ended generation, Microsoft-Decision-1.1 is built for high-frequency decision workflows inside enterprise software. It fits routing, classification, priority sorting, content validation, workflow governance and other tasks that require structured, bounded judgments. Instead of generating free-form long text, it selects valid outcomes within predefined rules supplied by developers, and returns calibrated confidence scores alongside each result. When input signals are ambiguous or insufficient, the model can abstain or flag cases for human review, lowering the risk of costly missteps in automated business pipelines.

Pricing on Microsoft Foundry varies by regional endpoint. The global standard tier charges $0.042 per million input tokens, with output tokens free of charge. Regional deployments carry slightly different rates: US data centers are priced at $0.046 per million input tokens, while EU and Asia Pacific endpoints cost $0.05 per million input tokens. This pricing structure makes it suitable for high-volume, low-latency decision jobs that would become economically impractical on general-purpose foundation models.

Benchmark results from Artificial Analysis Decision Index v2.1 show Microsoft-Decision-1.1 scores 68, a 6-point improvement from the original Microsoft-Decision-1. This score places it ahead of Jev 1.14, Google Gemini Decision 1 and OpenAI GPT-6 Luna (Decisions). It ranks eighth in the public decision-model leaderboard. For context, it beats Elephant 9B and Jev 1.14 by 2 points, while trailing d1.2 ultra by 1 point. The evaluation suite covers 102 separate benchmarks measuring knowledge, reasoning, classification, routing, agent tool use, preference judgment, safety and consistency, giving a multi-dimensional view of decision reliability.

Microsoft highlights two core performance gains in this release. The updated model delivers better intelligence and calibration, while its throughput is over twice that of the first version. It is optimized on a mid-sized mixed-expertise architecture based on Gemma 4.26B A4B, and runs across A100, H100, H200, GB200 and other mainstream GPU hardware. Scheduling optimizations help suppress per-inference overhead, which is critical for millisecond-scale decision triggers embedded directly into application logic.

Training combines supervised fine-tuning and reinforcement learning. The dataset includes synthetic and licensed samples spanning tens of thousands of decision task types, designed to improve generalization on unseen business rules. The model supports multimodal inputs, accepting images such as photographs and screenshots for visual decision workflows. Since launch, more than 4,000 customers have adopted it. Early adopters include Epic, Equinor, Atos Group and Raiffeisen Bank, using it for ticket routing, customer service triage, workflow automation, content tagging and classification.

Inside Microsoft’s own product stack, Microsoft-Decision-1.1 powers the routing logic for Copilot. Its judgment accuracy is comparable to large general-purpose models, but it runs 2.2 times faster and costs 4.4 times less. It also serves as the reasoning controller for Cowork, and feeds quality and risk assessments into Foundry evaluations.

For engineering teams running mixed-model fleets, switching between general-purpose LLMs and specialized decision models adds API routing complexity. 4sapi, an API gateway, helps normalize request formats and traffic management across different model providers.

The release underscores a clear industry shift: not every AI task needs a high-end generalist model. Many production automation needs only constrained, repeatable, fast decision outputs. Specialized decision models trade open-ended creativity for deterministic calibration, low latency and predictable cost. Teams can reserve powerful general models for deep analysis and creative generation, while offloading routing and classification jobs to lightweight decision models like Microsoft-Decision-1.1.

Even with strong benchmark results, developers should validate against internal business datasets. Benchmark performance does not always translate to rule-heavy, domain-specific workflows. The abstention feature is a key safety lever: teams can set confidence thresholds to pause automation and escalate uncertain cases to human operators, balancing speed and operational risk.

Tags:Microsoft Decision 1.1Microsoft FoundryMicrosoft AIAI Decision ModelsDecision Intelligence

Recommended reading

Explore more frontier insights and industry know-how.