Microsoft debuts Decision-1 to help AI agents choose actions
Microsoft has unveiled Decision-1, a compact scoring model that assigns calibrated probabilities to a fixed set of possible answers. Its intended uses include sorting requests, labeling items, ranking priorities, checking outputs, and controlling workflows, so applications and agents can decide whether to act, wait, or ask for human review. Microsoft tested it on many benchmark questions and says it runs faster than comparable options; internal trials include organizing customer feedback and helping evaluate AI responses during service incidents.
Microsoft built Decision-1 as a compact scorer rather than a generative chatbot. It was post-trained from Qwen3.5-9B to score answers in one pass, returning confidence values for choices such as yes/no, multiple-choice, ratings, or rubric grades through structured API calls. It is not intended for open-ended writing, dialogue, translation, or summaries.
The company tested it on 36 blind benchmarks with almost 150,000 questions. It reportedly ran 2.5 times faster than H2O-Lightning-4B v1.1 and roughly 35 times faster than GPT-6 Sol in median latency. It is available via Microsoft Foundry and OpenRouter, priced at $0.042 per million input tokens, with outputs free.
Decision-1 could affect developers, enterprises, and people interacting with AI agents. Faster, cheaper scoring may let services route requests, flag content, or escalate cases more consistently, potentially reducing delays and costs. But when probabilities guide whether a request is acted on, deferred, or sent to a person, errors or poorly chosen thresholds may disadvantage users. Human review and validation could remain important, especially in sensitive workflows.