Microsoft Launches Microsoft-Decision-1 for Low-Latency AI Decision-Scoring
Microsoft has introduced Microsoft-Decision-1, a specialized decision model post-trained on Qwen3.5-9B that delivers fast, calibrated scoring for agentic workflows.
Microsoft has officially released Microsoft-Decision-1, a specialized model designed specifically for fast decision-scoring and structured classification tasks. Rather than generating long-form natural language responses like traditional large language models, decision models are engineered to return structured outputs—such as probabilities over a predefined set of choices—that downstream software and automated agents can immediately execute.
Now available in Microsoft Foundry and via OpenRouter, Microsoft-Decision-1 is aimed at use cases such as model routing, incident prioritization, response verification, and agent workflow control. In an internal evaluation covering 36 blind benchmarks and nearly 150,000 questions, Microsoft reported that the model achieved the highest accuracy among tested systems while showing a median (P50) latency approximately 35 times faster than GPT-6 Sol and 2.5 times faster than the runner-up model, H2O-Lightning-4B v1.1.
Architecture, Calibration, and Benchmarks
To develop Microsoft-Decision-1, the team post-trained Qwen3.5–9B specifically for single-pass decision scoring. Microsoft also plans to rebase the system onto other foundations in the future, including models from OpenAI and Microsoft AI (MAI). Instead of generating unconstrained tokens, Microsoft-Decision-1 evaluates a fixed set of options—such as binary yes/no choices, multiple-choice lists, numerical rating scales, or rubric-based grading criteria—and returns calibrated probability scores via a structured API call.
Microsoft highlighted several engineering priorities addressed during the model's design:
- Robustness: Across tests where prompts were altered using eight different types of formatting noise, reordered choices, or paraphrased instructions, the model flipped its final decision on only 1.3% of perturbations, registering zero decision flips when options were paraphrased, reversed, or shuffled.
- Confidence Calibration: The output probabilities directly reflect predicted accuracy, allowing systems to establish automated confidence thresholds for deferring or triggering human review.
- Generalization: Tested against public benchmarks such as JevBench along with private evaluation suites, the model demonstrated strong performance across out-of-distribution, reasoning, and long-context tasks.
- Safety Screening: Evaluated against 5,250 adversarial requests across 11 safety benchmarks, the model consistently refused prompt injection and jailbreak attempts without over-filtering benign requests.
Internal Production Use Cases at Microsoft
Before releasing the model publicly, Microsoft deployed Microsoft-Decision-1 across several internal engineering divisions to measure real-world performance against larger general-purpose LLMs.
At Xbox Research, the team used the model to categorize more than 10,000 pieces of unstructured feedback collected from Steam, social media, and customer surveys into defined research themes. Microsoft reported that Microsoft-Decision-1 achieved classification quality comparable to GPT-6 Sol while operating more than 14 times faster and at 200 times lower cost.
Similarly, Microsoft's Copilot team utilized the model for automated quality control to grade chat and agentic responses. In those tests, Microsoft-Decision-1 delivered accuracy competitive with GPT5.6 Luna while operating 100 times faster. In incident response environments, on-call teams used the model to retrieve knowledge from logs, ticketing systems, and call transcripts, where it outpaced standard LLM retrieval pipelines. Additionally, in Microsoft Discovery's adaptive replanning setup for scientific workflows, the decision model produced scores that were 46 times more consistent than an LLM baseline at three times the evaluation speed.
What it means for developers
For developers building complex agentic pipelines, sequential decision-making represents a major performance bottleneck. Adding even 100 milliseconds across dozens of dependent evaluation steps can introduce seconds of lag. Microsoft-Decision-1 offers an alternative to routing lightweight evaluation, data validation, and intent detection through heavy conversational models.
The model is priced at $0.042 per million input tokens, with output tokens provided for free. This pricing structure makes it feasible to run high-volume screening operations—such as safety filtering, code scanning, rubric grading, and multi-model routing—without incurring large API bills.
In modern system architectures, developers frequently separate decision-making from generation: lightweight classification models handle routing and validation, while large multimodal models handle creative generation or deep synthesis. Developers looking to build and experiment with these diverse architectures can try top AI models cheaply through one API at https://apixoai.online to find the right balance between speed, cost, and output quality across their application stack.
As agentic workflows expand from theoretical prototypes into real-time production software, purpose-built decision models like Microsoft-Decision-1 provide teams with a specialized tool for deterministic, low-latency control flow.
Source: Introducing Microsoft-Decision-1: Fast AI for Real-Time Decisions — Towards AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
