Microsoft built a model that doesn't talk. That's the point.

Satya Nadella posted on October 9 about a new Microsoft model. The marketing hook — "fast decision-making AI" — sounds like every other AI launch. The actual product is different enough that I read it twice to make sure I wasn't misunderstanding.

Microsoft Decision-1 doesn't generate text. It takes a task description, a fixed list of options, and optional grading criteria, then returns a calibrated probability score for each option in one pass. No generation. No reasoning chain. Just: here are the choices, here's the best one, here's how confident I am.

This is the kind of model that slides under most builder radar because it doesn't do the thing that gets coverage.

What it actually is

Decision-1 is post-trained from Alibaba's Qwen3.5-9B, an open-weight 9-billion-parameter model. Microsoft didn't build this from scratch — they post-trained an existing open-weight base and optimized it for classification. That's either an embarrassing omission from the launch materials or a sensible engineering decision they're underselling. I think it's the latter.

The use cases Nadella called out: routing, classification, prioritization, verification, and workflow control. Microsoft's own Xbox Research team reportedly used it to categorize over 10,000 player reviews at GPT-6 Sol quality at 14× the speed. That sentence should land with anyone running any automated classification pipeline at scale.

Pricing: $0.042 per million input tokens, output free. The output-free part is load-bearing. Classification models don't generate many tokens — they return probabilities — so Microsoft can price by input only. For comparison, the cheapest tier of GPT-6 Luna is $0.20/M input and $0.80/M output. At any meaningful classification volume, the cost difference is not small.

Source spread

• Satya Nadella on X, Oct 9 — hype. Launch post. Covers routing, classification, prioritization, verification, workflow control. References Xbox Research and incident response as internal testers.
• Windows Report — hype. Covers the 35× speed headline without engaging with methodology.
• Windows Forum megathread — skeptic. Flags that all benchmark figures come from Microsoft's own evaluation, not third-party reproduction. Explicitly notes the Qwen3.5-9B base model wasn't prominently disclosed in launch materials.
• FourWeekMBA — builder. Clean breakdown of pricing and Foundry availability.

Pros & cons

What's real:

• The speed claim is structurally plausible. Decision-1 doesn't generate tokens — it scores a fixed option list in one pass. GPT-6 Sol has to generate a response. These are different compute profiles, and a large latency gap between them isn't surprising.
• The pricing is genuinely cheap for what it does. $0.042/M with free output changes the cost math for classification-heavy workflows.
• Availability in Microsoft Foundry on day one. This isn't a preview or a waitlist. You can test it now.

What deserves a side-eye:

• The benchmarks are company-reported. 83.5% average accuracy across 36 suites with 147,137 questions sounds rigorous. It might be. But these are benchmarks Microsoft selected, run by Microsoft. No independent reproduction yet.
• The 35× latency comparison is against GPT-6 Sol specifically — the heaviest model in OpenAI's lineup. A classification model should be faster than the biggest frontier LLM. The comparison flatters the headline.
• The Qwen3.5-9B foundation. Building on Alibaba's open-weight model while separately benefiting from US export controls on Chinese AI hardware is a tension worth naming. It's not wrong — open weights are open weights — but it's there.

What builders need to know

Further reading

• Satya Nadella on X — Launch announcement — primary source
• MarkTechPost — Technical overview and Qwen3.5-9B foundation
• Windows Forum megathread — Benchmarks and pricing caveats
• FourWeekMBA — Pricing breakdown
• Futunn — Xbox Research internal use case