Microsoft-Decision-1

Calibrated confidence means that when the system gives a score of 0.9, the probability of being correct should be close to 90%.

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

Microsoft CEO Nadella stated that Microsoft is testing the model across various areas of the company, covering scenarios such as incident response, quality control, and scientific discovery

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.

The Xbox research team used Decision-1 to categorize over 10,000 pieces of feedback, achieving performance comparable to GPT-6 Sol but more than 14 times faster and 200 times cheaper.

Decision-1 is live on Microsoft Foundry, priced at $0.042 per million input tokens, with output being free.

d1-3B's question types include noul (returns P(yes)), choice (returns the selected option, confidence, and the probability of each option), and score (returns the expected rating, confidence, and probability distribution on a 2–10 ordinal scale).

Microsoft-Decision-1 has gone live on the Microsoft Foundry and OpenRouter platforms

Microsoft-Decision-1 was released on October 9, with an input price of $0.042 / 1M tokens, free output, and undisclosed weights and context.

A Microsoft-Decision-1 request is not a chat completion; it requires passing in state (a string or JSON) plus named questions.

The model ID for Microsoft-Decision-1 on OpenRouter is microsoft/microsoft-decision-1, and the official documentation states that the chat-completions SDK cannot be used and the Decisions / System One interface must be used instead.

OpenRouter lists Microsoft-Decision-1's context window as 32,768 tokens, with text-only input and structured decision output, and no support for images or audio.

In Copilot quality scoring, Microsoft-Decision-1 is on par with GPT-5.6 Luna and about 100 times faster (Microsoft's statement).

The endpoint path for Microsoft-Decision-1 on Foundry is <your-foundry-resource-endpoint>/providers/microsoft/v1/systemone, and the model field in the request is filled with the deployment name rather than the model string.

According to Achint Srivastava's blog, Microsoft-Decision-1 had the highest accuracy in Microsoft's comparison of 36 benchmarks, with about 150,000 questions kept blind during training.

The choice type always selects one option, and if no abstention option is provided, the model is forced to give an answer even when evidence is insufficient; the Learn documentation recommends adding an abstention option when a choice should not be forced.

The model input cost for 100,000 requests × 2,000 billed input tokens is about $8.40 (excluding retries, retrieval, moderation, and platform fees).

Under eight kinds of input perturbation, Microsoft-Decision-1's average decision flip rate is only 1.3%; when option descriptions are reworded or option order is reversed/shuffled, there are zero flips (vendor data).

The balance between speed and quality in specialized decision models attempts to make trade-offs between two metrics that often conflict with each other.

Embedding a decision model into incident response workflows can automate the first two steps of the 'alert triage → root cause localization → remediation matching' pipeline.