Microsoft Launches Decision-1: A Fast Decision-Scoring Model Built on Qwen3.5-9B

Microsoft has released a new model focused on decision scoring called Microsoft-Decision-1, purpose-built for scenarios such as routing, classification, prioritization, outcome verification, and workflow control. The model was announced by Achint Srivastava, Corporate Vice President of Software Engineering in the Office of the CTO, and is currently available in public preview through Microsoft Foundry and OpenRouter.

Unlike general-purpose large language models that generate free-form text, Decision-1 takes a state description and a fixed set of candidate answers, then returns a calibrated probability score for each option. This structured output is designed directly for software systems that need to act on results immediately — not for humans reading long-form responses.

How the Model Works

Microsoft built Decision-1 through post-training on top of Alibaba's open-source Qwen3.5-9B weights, specializing it for single-pass decision scoring tasks. The company says it plans to port the model to other base models in the future, including Microsoft's own MAI and OpenAI's models. The current version supports yes/no questions, multiple-choice questions, rating tasks, and rubric-based scoring of AI responses or agent actions.

The model accepts plain-text input with a context window of 32,768 tokens and returns output in JSON format. Pricing is $0.042 per million input tokens, with output tokens completely free. Notably, Microsoft has not released the weights — the model is available exclusively as a hosted API.

Microsoft highlighted five design goals: low latency (P50 of approximately 85 milliseconds in internal testing), generalization across unseen benchmarks, robustness to paraphrasing and option reordering, probability calibration quality, and safety filtering that blocks harmful requests while preserving utility.

---

Post-training refers to the process of further fine-tuning a pre-trained large model toward specific tasks or behavioral objectives — typically involving supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF). Compared to training from scratch, post-training preserves the base model's language understanding while dramatically narrowing its output space for a given task. That's exactly what Decision-1 does: it uses Qwen3-9B's general language capabilities as a foundation, then uses post-training to lock the output to structured probability scores rather than open-ended text generation. Probability calibration measures whether a model's output confidence aligns with its actual accuracy — if a model assigns 80% probability to an option, that option should be correct approximately 80% of the time in a statistical sense. Poorly calibrated models can be overconfident or overly conservative, causing downstream systems to make incorrect decisions based on distorted confidence scores. This is why Microsoft lists its 92.2 calibration score as a standalone core performance metric.

---

Benchmark Results

In Microsoft's internal evaluation covering 36 benchmarks and nearly 150,000 questions (all withheld from training), Decision-1 ranked first with an average accuracy of 83.5%. It outperformed Jev 1.13.0 (82.3%), Quyet-1.0-Large (81.9%), GPT-6 Luna Decisions (79.4%), and several other specialized decision models.

Latency is a standout characteristic: the report claims Decision-1 is approximately 2.5× faster than the second-ranked decision model, and roughly 35× faster than GPT-6 Sol on the same tasks. Its calibration score is 92.2 out of 100. Robustness testing showed an average of only 1.3% decision change across various perturbations, with zero decision flips when option descriptions were paraphrased or options were reordered.

It's worth noting that all of these figures come from Microsoft's own reporting. At the time of launch, no independent third-party benchmarks exist, meaning real-world performance still awaits external validation.

Real-World Use Inside Microsoft

Microsoft has already deployed Decision-1 across several internal use cases:

• The Xbox research team used it to categorize over 10,000 open-ended feedback responses into researcher-defined themes, finding it comparable to GPT-6 Sol in quality but more than 14× faster and 200× cheaper.
• The Copilot team uses it to measure the quality of chat and agent responses, reporting competitive performance with GPT-5.6 Luna at roughly 100× the speed.
• On-call engineers are testing it for retrieving relevant knowledge during real-time incident response.
• Microsoft Discovery uses it for adaptive replanning in scientific experiments, reporting higher consistency and nearly 4× overall speed improvement compared to LLM-based approaches.

These examples clearly illustrate the model's positioning: it's a fast, reliable decision layer embedded within larger agent or workflow systems — not a replacement for generative models.

What This Means for Developers and Enterprises

This category of decision model reflects a broader trend toward modular AI architectures. Not every step in an agent or application requires open-ended generation — routing requests to the right tool, identifying user intent, validating intermediate results, or scoring agent actions can often be handled more efficiently by a dedicated scoring model.

For developers building on Microsoft Foundry or using OpenRouter, Decision-1 offers a low-cost, low-latency option for these structured tasks. With free output tokens and extremely low latency, it can be inserted into multi-step pipelines without consuming a significant share of the cost or latency budget. The closed-weight, Azure-hosted nature may appeal to organizations that prefer managed models over self-hosted open-source alternatives.

That said, since performance relies on Microsoft's own benchmarks, teams should validate results against their own specific decision distributions before moving to production. The model also currently supports only text — no multimodal inputs — which is a functional limitation compared to some competing decision APIs that accept image inputs.

---

The core idea behind modular AI architecture is to decompose complex agent pipelines into single-purpose, independently optimizable components rather than relying on one large general-purpose model to handle every step. In a typical multi-agent system, a router decides which tool to call, a verifier checks whether intermediate results meet expectations, and a scorer selects the best candidate action — steps that require minimal generative capability but demand high speed and determinism. Offloading these tasks to lightweight specialized models lets expensive generative models focus their compute on steps that genuinely require creative output, while reducing per-call latency to support more complex multi-turn decision chains. Decision-1 is positioned as the "decision layer" in this kind of architecture, and its free output pricing is clearly designed with high-frequency pipeline call patterns in mind.

---

Why This Matters

The release of Microsoft-Decision-1 further demonstrates that progress at the AI frontier is no longer solely about building larger generative models. Specialized architectures optimized for speed, calibration, and structured output are maturing into practical building blocks. For organizations already deeply invested in the Microsoft ecosystem, it offers an immediately usable tool to make agent workflows faster and more reliable. For the broader industry, it signals that decision-scoring models are emerging as a distinct and commercially viable category, developing in parallel with traditional LLMs.