Decision-1

Decision models focus on answering yes/no judgments, multiple-choice questions, and probability-scored rating-type problems.

Microsoft launched Microsoft-Decision-1 on Product Hunt, a decision scoring model specifically designed for decision-making.

Microsoft reported that Decision-1 achieved the highest accuracy in a comparison across 36 benchmarks covering nearly 150,000 questions, and claimed its P50 latency is approximately 35 times faster than GPT-6 Sol.

d1's decision calls appear in AI Gateway logs and are counted in budget statistics, allowing reuse of AI Gateway's cost attribution, call auditing, and quota control processes.

Microsoft's Decision-1 was announced by Achint Srivastava, VP of the Office of the CTO, and there are plans to later redo the base on Microsoft AI and OpenAI models.

In AIOps root cause localization practice, it is recommended to use Decision-1 to rank troubleshooting directions rather than directly naming fault types, and to retain a human review step for major decisions.

Decision-1's threshold of 0.5 is a bad default value; the threshold must be optimized on a calibration set, and the best threshold in this paper's HDFS task is 0.15.

In HDFS anomaly detection, Decision-1's raw output probabilities correlate 0.43 with event count and only 0.14 with labels; the model overreacts to lexical cues and log volume, and misjudges missing events as healthy.

Microsoft introduced the Decision-1 model, which does not generate natural language and only outputs probability values: yes/no judgments output noul, multiple-choice judgments output choice, and ordered ratings output score, consuming about 1 output token per call.

In experiments on the Loghub HDFS_v1 dataset, Decision-1 initially used a simple prompt and reached only 49% accuracy (AUC 0.68), lower than the 71% rule-based baseline.

In HDFS anomaly detection, after using structured questioning (computing lifecycle_summary and changing the question to 'is the lifecycle complete'), Decision-1 reached 86% calibrated accuracy on 1200 test blocks (threshold 0.15), with AUC reaching 0.964.

On the RCAEval RE2-OB microservice root cause localization dataset, Decision-1 correctly identified the root cause service 85 times out of 90 cases, with 94% accuracy, exceeding the 82% of the 'maximum metric deviation' rule-based baseline (McNemar p=0.013).

On the RCAEval RE2-OB fault type identification task, Decision-1 achieved only 51% accuracy, correctly identifying 13 of 15 cases each for CPU and latency faults, but 0 correct for packet loss faults and only 2 correct for socket exhaustion.

Decision-1 has an average latency of 240ms, with 1380 calls costing about $0.015; gpt-4.1-nano has an average latency of 728ms and a cost of about $0.05-0.07; gpt-6.1-sol has an average latency of 2052ms and a cost of about $0.56.

At the same call volume, Decision-1 is about two orders of magnitude cheaper than reasoning models.

Microsoft's Decision-1 is a structured classification model that outputs a probability distribution rather than natural language, and supports three SDK integration methods: AI SDK, OpenAI-compatible, and TypeSafe.

Decision-1 and Clef-omni require first aligning with existing latency, token cost, and context window baselines before deciding whether to replace, and it is recommended to evaluate first.

Decision-1 can be called in three ways: the AI SDK's decision API, the OpenAI-compatible Decisions API, and the TypeSafe-compatible API.

Decision-1 is an AI model launched by Microsoft focused on structured decisions, whose outputs are probabilities, options and scores, rather than free text.

Developers can bind multiple questions to the same input and complete multiple judgments in a single request.