Microsoft introduces Decision-1, a model for fast structured choices

Microsoft has introduced Microsoft-Decision-1, a model designed to make bounded choices for software instead of writing an open-ended answer. Microsoft’s announcement is dated October 9, 2026; the page metadata records publication on October 10 and an update later that day. The model is available in public preview through Microsoft Foundry and OpenRouter.

The distinction is practical. A conventional language model might explain several options, while Decision-1 is asked to select from fixed answers, return yes or no, assign a rating or apply a rubric. Microsoft presents it as a post-trained Qwen3.5-9B model for tasks such as routing requests to another model, classifying data, judging an AI response, triaging incidents and checking whether an agent action should proceed. That makes it a possible control layer around an agent rather than a general-purpose chatbot.

Microsoft reports the highest accuracy across 36 internal benchmarks and nearly 150,000 blind questions. The company says Decision-1 was 2.5 times faster than H2O-Lightning-4B, the runner-up in its comparison, and 35 times faster than GPT-6 Sol. Microsoft also reports only 1.3% decision changes when prompts are lightly perturbed. Its safety evaluation covered 5,250 requests across 11 benchmarks. Those figures are vendor results, so they describe Microsoft’s test design and should not be treated as an independent ranking.

Early outside evidence is more qualified. A Calibrated Agents comparison found no universal winner and measured Decision-1 as slower in its own test, while another comparison noted that the model had not yet appeared on third-party model boards as of October 10. The product’s low listed price—$0.042 per million input tokens with free output—could still make repeated checks affordable, but teams need to measure accuracy, calibration and false approvals on their own data.

The launch matters because reliable fixed-choice decisions are a missing piece in many agent systems. Cheap, fast checks can route work, request human review or block a risky action. They cannot replace explanations or judgment, however. Microsoft-Decision-1 is best understood as a specialized decision component whose value will depend on independent testing and careful safeguards.