On October 9, 2026, Microsoft introduced Microsoft-Decision-1, an AI model that does not write answers: it takes a piece of text and a fixed list of options and returns a probability for each one. Developers can use it in Microsoft Foundry and on OpenRouter. It matters to anyone sorting messages, filtering content or policing a chatbot without paying for a large model.
What we know
- What it is: Microsoft calls it a "decision model." According to its October 9 post, it is built for routing, classification, prioritization and verification inside applications and agents.
- How it answers: the Microsoft Learn documentation describes three question types: yes or no, pick one option, and score on a scale. It returns numbers, with no written rationale.
- What it is built on: Microsoft says it post-trained the Qwen3.5-9B model and will soon rebase it on other models, including its own (MAI) and OpenAI's.
- Price: $0.042 per million input tokens. Output tokens are free, according to the post.
- Where to get it: in Microsoft Foundry, which requires an Azure subscription with a payment method according to the documentation, and on OpenRouter, which lists it from October 9 with a 33,000-token context window and Azure as the only provider.
- What Microsoft measured: the highest accuracy in its own 36-benchmark comparison covering nearly 150,000 questions, and median latency 35 times lower than GPT-6 Sol. The post was updated after publication and now names H2O-Lightning-4B as the runner-up, 2.5 times slower.
- What Microsoft admits: the documentation warns that scores can shift with the wording or order of the options, that calibration is strongest on familiar tasks, and that the scores should not be the sole basis for decisions about individuals.
What changes and what doesn't
The cost of a very common task drops. With a chat model, sorting a message or checking whether a text follows a rule also pays for the text of the reply. Here the reply is a number, and it is free.
The need to test does not go away. The performance figures are Microsoft's, and no third party has measured them yet. Neither the post nor the documentation says which languages it handles best; the post only lists multilingual tasks among its benchmarks. And the OpenRouter listing notes that the model weights are updated continually while the API stays the same: what works today may behave differently in a few weeks.
It is not a chatbot either. The model does not chat, translate or summarize. It handles the step in between: deciding what to do with a piece of text before a person or another model acts on it.
How to tell if it affects you
- List the repetitive decisions on your site. Contact-form messages that need routing to sales, support or spam; blog comments waiting for approval; chatbot conversations that should go to a human.
- Do the math. At 10,000 messages a month of about 500 tokens each, counting the question and the options, you send 5 million tokens: roughly $0.21 at the published price. This is an estimate made for this article and leaves out any other Azure or OpenRouter costs.
- Test it on 50 real cases. Before letting it decide alone, run 50 messages you have already labeled by hand, in the languages your customers use, and compare. If it misses more often than you would accept from a new hire, do not automate it. Rerun the test from time to time, because the model keeps changing.
- Set a threshold. Because it returns probabilities, you can act only above, say, 90% and send the rest to human review. The documentation asks you to validate that threshold against labeled examples.
- Watch your chatbot. The documentation's own example asks whether a reply promises a refund. The same idea can stop replies that promise prices, delivery times or guarantees your business does not offer.
- Don't decide about people. For screening candidates, approving credit or similar calls, Microsoft itself advises against relying on this score alone.
- If you don't code, ask whoever runs your site or chatbot whether they use a chat model to classify things and what it costs. That is where this kind of model can cut the bill.
Related: Claude Haiku 5.5: Up to 90% Cheaper, With a 100K-Token Catch
Sources
- Microsoft (Command Line), "Introducing Microsoft-Decision-1, our model for fast decision-making," Achint Srivastava, October 9, 2026
- Microsoft Learn, "Deploy and use Microsoft-Decision-1 in Microsoft Foundry"
- OpenRouter, Microsoft-Decision-1 model page
Updates: this note will add independent measurements of the model and any price or base-model change Microsoft announces.