Microsoft's Decision-1 makes the yes/no call cheap

If your agent asks a big model "should I do this?" all day, Microsoft wants that question answered by something much smaller and much cheaper.

On Oct. 9 Microsoft published Microsoft-Decision-1, a model that does one job: you hand it a fixed set of options and it returns a probability for each. Yes or no. A, B or C. A score against a rubric.

Input costs $0.042 per million tokens. Output is free.

A small model built to make one call

Decision-1 is Qwen3.5-9B, post-trained by Microsoft to score decisions in a single pass, according to the post by Achint Srivastava, a vice president in Microsoft's Office of the CTO. Microsoft says it will soon rebase the model on others, including its own MAI models and OpenAI's.

The pitch is structured output software can act on. Microsoft lists routing, classification, prioritization, verification and grading of AI responses and agent actions. The probabilities are part of the API. Microsoft's stated bar: "a 90% prediction should be right about nine times out of 10 on representative cases."

It is on Microsoft Foundry and OpenRouter now, with documentation on Microsoft Learn.

The benchmark is Microsoft's. The price is real.

Microsoft says Decision-1 had the highest accuracy in its own 36-benchmark comparison, nearly 150,000 questions held out of training. It says the model is 2.5 times faster than the runner-up, H2O-Lightning-4B, and about 35 times faster than GPT-6 Sol at median latency. Those are Microsoft's tests, run on its chosen rivals. Treat them as a claim until someone else reruns them.

The robustness numbers are more interesting. Across eight kinds of rewording, Microsoft says decisions flip 1.3% of the time on average, and never when the options are paraphrased, reversed or shuffled. If you have ever watched a large model change its answer because you swapped A and B, you know why that matters.

Microsoft also reports internal results. Xbox Research labeled more than 10,000 open-ended feedback items with quality it calls competitive with GPT-6 Sol, at more than 14 times the speed and about 200 times lower cost, the post says.

Here's the thing. Decision-1 is plumbing, and plumbing is where agent bills go to die.

Where it fits in what you build

If you run an AI agent, a lot of its spend goes on small judgments: is this ticket urgent, does this answer pass, which model should take this request, is this action safe to run. Send those to the same frontier model that does the real work, and you pay frontier prices for a yes or a no.

Decision-1 is the argument for splitting that out. A cheap scorer at the gate, the big model only where it earns its cost. Because the output is a probability, you can also set a threshold: act above 0.9, ask a human below it.

The category is getting crowded. TopFive covered Cloudflare and TypeSafe AI's rival decision models on Saturday. Microsoft entering it tells you the big platforms think this layer is worth owning.

Two caveats from the post itself. It does not state a license or say whether weights are downloadable. And it is a 9B model: anything that needs real reasoning still belongs to a bigger one.

Watch the rebase and the independent scores

Watch for the rebase. A version built on an OpenAI model, sold inside Microsoft's cloud, would put this layer in front of a lot of enterprise agents by default. Watch independent scores too, since every number above is Microsoft's.

Then price your own gatekeeping calls. Count how many of your agent's model calls are really a yes or a no.