Model Releases · OCTOBER 11, 2026
Microsoft-Decision-1 ships at $0.042/M on a 9B Qwen base, with output tokens free
Microsoft's first decision-scoring model returns calibrated probabilities over fixed answer options in a single pass, posting 83.5% average accuracy across 36 blind benchmarks and an 85 ms p50 latency on Microsoft's own tests.
Microsoft announced Microsoft-Decision-1 on October 9, 2026, pricing a purpose-built decision-scoring model at $0.042 per million input tokens with output tokens free. The model is post-trained from Alibaba's open-weight Qwen3.5-9B base, which is already an uncomfortable detail given Microsoft's typical narrative management around its own stack.
Decision-1 doesn't do chat. It takes a question and a fixed option set (yes/no, multiple choice, a rating ladder) and returns a calibrated probability per option as JSON in a single pass, over a 32,768-token context window. The model card is explicit that it isn't meant for open-ended Q&A, translation, summarization, or sole automated calls on credit, employment, housing, healthcare, or legal rights.
The pitch is cost and latency. On Microsoft's own 36-benchmark suite of 147,137 blind questions, Decision-1 posts 83.5% average accuracy and a 92.2 calibration score, at an 85 ms p50 and 125 ms p95. Microsoft claims it runs 4.5× faster than runner-up Quyet-1.0-Large, 2.8× faster than TypeSafe's Jev (which carries the same $0.042/M input price), and 35× faster than GPT-6 Sol. Quyet-1.0-Large edges Decision-1 on calibration at 93.1.
As Microsoft frames the moment: "Now that agentic AI is a reality, we've seen that cost plays a major role in how people decide to use AI."
The internal deployment numbers are the softer evidence. Microsoft says Xbox Research sorted more than 10,000 feedback items at 14× the speed and 200× the cost efficiency of GPT-6 Sol; Copilot quality control ran 100× the speed of GPT-5.6 Luna; Microsoft Discovery got 46× more consistent scoring and a 3× speed improvement. These are Microsoft's own measurements, not independent benchmarks, and Decision-1 doesn't appear on the public JevBench leaderboard. H2O.ai's model card contests the latency figures Microsoft used for H2O-Lightning-4B, claiming its own measured median is 29 ms.
Distribution is where the strategy shows. Decision-1 is live in Microsoft Foundry's Direct-from-Azure GA tier and on OpenRouter as microsoft/microsoft-decision-1, routed through a Decisions API rather than the chat-completions endpoint. The Register reports Microsoft plans to rebase future versions on MAI and OpenAI backbones without providing a timeline, which quietly concedes that the Qwen foundation is a bridge, not a destination.
The structural read is simpler than the benchmark theater. Microsoft shipped a sub-cent streaming tier Microsoft shipped last week, watched the Gemini 4 Argon reset and the GPT-6.1 Sol made at the frontier, and concluded that the growth market isn't the frontier. It's the millions of routing calls per hour that frontier tokens were quietly subsidizing.