Image: Microsoft

Most of what AI agents do all day isn't writing essays. It's making small calls: is this email spam, which team gets this ticket, should the agent click "buy" or stop and ask a human. Microsoft's new Microsoft-Decision-1 is built only for those calls, and it's priced to make running thousands of them almost free.

Introducing Microsoft-Decision-1

Our model for fast decision-making that delivers top performance in latency and quality on structured decision tasks to outperform both LLMs and other decision models.
https://t.co/jjtqKq34kV

— Microsoft Developer (@msdev) October 9, 2026

Microsoft announced the model on its Command Line blog. It's available now in Microsoft Foundry and through OpenRouter. Instead of writing a reply, it takes a question and a fixed list of answers, then returns a probability for each option. It handles yes/no questions, multiple choice, ratings and rubric-based grading of other AI models' answers or agent actions.

That probability is the useful part. An app can act on its own when the model is 95% sure and send the 60% cases to a person. Microsoft says its scores are calibrated, meaning a 90% answer should be right about nine times out of 10.

Pricing is where this gets interesting. Input costs $0.042 per million tokens and output is free. OpenAI's rival Decisions API, which recently opened to all developers, charges $0.10 per million input tokens. Sorting a million 500-token support tickets would cost about $21 with Microsoft's model versus roughly $50 with OpenAI's, before any extra fees.

Microsoft also makes big speed claims. It says the model is about 35 times faster than GPT-6 Sol at typical latency and 4.5 times faster than Quyet-1.0-Large, the runner-up in its testing. It also says it scored the highest accuracy across 36 benchmarks with nearly 150,000 questions kept out of training.

Those numbers deserve some care. They're Microsoft's own tests, and the headline speed comparison is against GPT-6 Sol, a full general-purpose model that was never designed for quick classification. OpenAI's own decision model, GPT-6 Luna Decisions, is on Microsoft's list of benchmarked models, but the blog doesn't give a direct speed or price comparison with it.

There's another detail worth knowing. Microsoft-Decision-1 isn't built from scratch. Microsoft says it took Qwen3.5-9B, a small open model from Alibaba's Qwen team, and trained it further for single-pass scoring. The company plans to rebase it on other models soon, including its own MAI models and OpenAI's. For businesses with rules about Chinese-made models, that origin may matter more than the benchmark scores.

The internal results give a sense of where it fits. Microsoft says Xbox researchers used it to sort more than 10,000 pieces of player feedback into themes, matching GPT-6 Sol on quality while running over 14 times faster at a cost 200 times lower. The Copilot team used it to check answer quality and found it about 100 times faster than GPT5.6 Luna.

Microsoft also tested how stable its answers are. When the same request was reworded or the options were shuffled, the model changed its decision 1.3% of the time on average, according to the company.

So who benefits? Mostly developers building agents, support tools, content filters and data-labeling pipelines, where the same small judgment runs thousands of times. Regular users won't pick it from a menu, but they may notice apps that respond faster and route requests better.

The limits are clear. It can't write text, explain its reasoning or call tools, and it only picks from options you provide. Nobody outside Microsoft has published independent benchmarks yet, and Microsoft hasn't said how the price will change once the model moves to its MAI or OpenAI base. The real test will be whether developers see the same speed and accuracy on their own data.