Microsoft launches Microsoft-Decision-1: a Qwen3.5-9B-based decision-scoring model at $0.042/M input tokens, free output, on Foundry and OpenRouter

Microsoft-Decision-1 is Microsoft's first in-house decision model, announced Oct 9, 2026 by Satya Nadella and detailed on Microsoft's Command Line blog. It is post-trained from Qwen3.5-9B for single-pass scoring: given a fixed set of options it returns a calibrated probability per option (yes/no, multiple choice, ratings, rubric grading of AI responses and agent actions), rather than generating text. Microsoft says it will later rebase the model on MAI and OpenAI models while keeping the API shape stable.

Vendor-reported results (no independent replication yet): highest accuracy across 36 benchmarks held out from training (~150K questions), 2.5x faster than runner-up H2O-Lightning-4B v1.1 and ~35x lower P50 latency than GPT-6 Sol, an average 1.3% decision-flip rate across eight perturbation types, and safety testing on 5,250 requests across 11 benchmarks. Internally, Xbox Research labeled 10K+ feedback items at quality near GPT-6 Sol while running 14x+ faster and ~200x cheaper.

It is available on Microsoft Foundry and OpenRouter at $0.042 per million input tokens with free output and a 32,768-token context. On OpenRouter it runs on the Decisions API (POST /api/alpha/decisions) instead of the chat completions endpoint. Good fits: agent step gating (continue/stop/retry/hand off), model routing, ticket triage, data labeling and AI judging. It is not meant for open-ended generation.