Microsoft-Decision-1 is a small, purpose-built model that scores structured choices rather than generating free-form text, aimed at developers who need deterministic calls inside agent pipelines, routers, and classifiers. The model targets developers building routing, classification, prioritization, verification, workflow control, agent controls, data labeling, AI judging, search relevance, safety screening, computer use, robotics, and scientific discovery systems. Announced by Satya Nadella and available now in Microsoft Foundry and due on OpenRouter soon , it lands as teams building agentic software increasingly need a cheap, fast control layer that sits between a large reasoning model and the actions it triggers.
Rather than generating open-ended text, Microsoft-Decision-1 takes a fixed set of options and returns a calibrated probability for each through a structured API call. It supports yes/no, multiple-choice and rating options, along with rubric-based grading of AI responses and agent actions. Microsoft post-trained Qwen3.5-9B for single-pass scoring and plans to rebase the system on other models, including Microsoft AI and OpenAI technology. Microsoft reports that across eight perturbations of equivalent requests, its decision changed 1.3% of the time on average, with no flips when option descriptions were paraphrased or options were reversed or shuffled , which matters for anyone who has watched an LLM-based judge contradict itself on reruns.
Internal use already covers meaningful ground. Xbox Research used it to label more than 10,000 feedback items and reported quality competitive with GPT-6 Sol while running over 14 times faster and costing 200 times less. The Copilot team found it competitive with GPT-5.6 Luna and 100 times faster. Pricing starts at $0.042 per million input tokens, with output tokens free. Microsoft is positioning the model as a low-cost control layer for agentic systems, where confidence scores can determine whether software acts, defers, retries, escalates or hands work to a model, tool or person.
Comments (7)
Sign in to join the discussion.
P
Priya Tandon2d ago
Naming it Microsoft-Decision-1 feels like it was decided by a model optimized for latency over quality. Try: Decider.
K
Kwame Brooks2d ago
The tweet getting cut off at 'from incident' is sending me. Decision-1 apparently decided we didn't need the rest of the sentence.
B
Benoît Larsson2d ago
Hey Satya, happy to chat on background about which internal incident response flows are running on this today. Even directionally would be great.
L
Linda Voss2d ago
Before I route this to procurement: SOC2 Type II, FedRAMP High, data residency in EU, and does it ship with the standard Azure DPA? Also need SSO on day one, not day ninety.
M
Mitsuru K.2d ago
Benchmarks on 'structured decision tasks' is doing an enormous amount of work in that sentence. Which tasks, whose structure, and compared to which LLMs at what quantization.
V
Valentina Gomez2d ago
Curious what the inference cost per decision looks like at Microsoft-internal scale. If this replaces GPT-class calls for routing, the margin story writes itself.
S
shanice2d ago
A fast small model for routing and triage is genuinely the unsexy win of the year. Everyone's shipping 2T param behemoths to decide if a ticket is spam.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Microsoft
Our model for fast decision-making
Our model for fast decision-making
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Microsoft-Decision-1 is a small, purpose-built model that scores structured choices rather than generating free-form text, aimed at developers who need deterministic calls inside agent pipelines, routers, and classifiers. The model targets developers building routing, classification, prioritization, verification, workflow control, agent controls, data labeling, AI judging, search relevance, safety screening, computer use, robotics, and scientific discovery systems. Announced by Satya Nadella and available now in Microsoft Foundry and due on OpenRouter soon , it lands as teams building agentic software increasingly need a cheap, fast control layer that sits between a large reasoning model and the actions it triggers. Rather than generating open-ended text, Microsoft-Decision-1 takes a fixed set of options and returns a calibrated probability for each through a structured API call. It supports yes/no, multiple-choice and rating options, along with rubric-based grading of AI responses and agent actions. Microsoft post-trained Qwen3.5-9B for single-pass scoring and plans to rebase the system on other models, including Microsoft AI and OpenAI technology. Microsoft reports that across eight perturbations of equivalent requests, its decision changed 1.3% of the time on average, with no flips when option descriptions were paraphrased or options were reversed or shuffled , which matters for anyone who has watched an LLM-based judge contradict itself on reruns. Internal use already covers meaningful ground. Xbox Research used it to label more than 10,000 feedback items and reported quality competitive with GPT-6 Sol while running over 14 times faster and costing 200 times less. The Copilot team found it competitive with GPT-5.6 Luna and 100 times faster. Pricing starts at $0.042 per million input tokens, with output tokens free. Microsoft is positioning the model as a low-cost control layer for agentic systems, where confidence scores can determine whether software acts, defers, retries, escalates or hands work to a model, tool or person.
Comments (7)
Naming it Microsoft-Decision-1 feels like it was decided by a model optimized for latency over quality. Try: Decider.
The tweet getting cut off at 'from incident' is sending me. Decision-1 apparently decided we didn't need the rest of the sentence.
Hey Satya, happy to chat on background about which internal incident response flows are running on this today. Even directionally would be great.
Before I route this to procurement: SOC2 Type II, FedRAMP High, data residency in EU, and does it ship with the standard Azure DPA? Also need SSO on day one, not day ninety.
Benchmarks on 'structured decision tasks' is doing an enormous amount of work in that sentence. Which tasks, whose structure, and compared to which LLMs at what quantization.
Curious what the inference cost per decision looks like at Microsoft-internal scale. If this replaces GPT-class calls for routing, the margin story writes itself.
A fast small model for routing and triage is genuinely the unsexy win of the year. Everyone's shipping 2T param behemoths to decide if a ticket is spam.