Microsoft-Decision-1: $0.042 per million, output is free

Microsoft has its own decision model now. Microsoft-Decision-1 does not write text. You give it a fixed set of options, and it returns a probability for each one in a single pass. Input costs $0.042 per million tokens. Output is free. For the routing, tagging and yes/no checks that fill most business automations, that is close to zero. The benchmark numbers are Microsoft's own, so we test before we switch.

What actually happened

Microsoft announced the model on October 9 in a post by Achint Srivastava, a VP in the Office of the CTO. A companion post on the Foundry blog covers the developer side. The key facts:

• What it does: routing, classification, prioritization, verification and workflow control. It handles yes/no, multiple-choice and rating questions, and it can grade AI responses and agent actions against a rubric.
• Base model: Qwen3.5-9B, post-trained for decision scoring. Microsoft says it will soon rebase the model on other models, including MAI and OpenAI models.
• Price: $0.042 per million input tokens, free output.
• Where: Microsoft Foundry and OpenRouter.
• Speed: Microsoft says median latency is about 35x faster than GPT-6 Sol.
• Stability: Microsoft perturbed the same request in eight ways. The decision changed on 1.3% of perturbations on average, and did not change when options were shuffled, reversed or paraphrased.

Microsoft also claims the highest accuracy across 36 benchmarks with almost 150,000 questions. The post does not name those benchmarks or give per-benchmark scores. Calibration (a 90% score that is right nine times in ten) is stated as a goal, without measured numbers. The Decoder puts the release in context: this is a crowded category now.

Why it matters for your business: decision models do the boring 80%

Look at a typical automation. Which inbox does this email go to? Is this order a fraud risk? Is this review a complaint? Did the agent's reply follow policy? None of those need a model that writes paragraphs. They need a model that picks one option and tells you how sure it is.

If you send those calls to a general LLM today, you pay for output tokens and you wait for them. A decision model removes both costs. It also gives you a number you can set a threshold on: auto-route above 0.9, send to a human below it.

What we would do this week:

Pull 200 real examples from one decision you already make, with the correct answers.

Run them through Decision-1 and your current model. Compare accuracy, latency and cost on your data, not Microsoft's.

Check the scores. If items scored 0.9 are right far less than 90% of the time, the threshold is not safe yet.

Keep the call portable. Microsoft is not the only vendor in this race; Perplexity sells one too. Put the model behind one routing function so the next price cut is a config change.

Sources: Microsoft Command Line, Microsoft Foundry Blog, The Decoder.