Microsoft has priced its new Decision-1 model at $0.042 per million input tokens and charges nothing for output. A price that low shows how quickly this category has stopped being a novelty and become a commodity.
Decision models are designed for quick, structured calls rather than long-form answers. Microsoft has now joined the field, but it is arriving late. OpenAI and Cloudflare already offer their own versions, and both came after Jev, the startup that launched the trend.
What Decision-1 is meant to handle
According to Microsoft, Decision-1 is built for classifications, evaluations and routing decisions. The company added that the model has “the potential” to “guide and control agents through complex environments.”
That second claim comes with a qualifier. Microsoft is describing what the model might do with agents, not what it has actually shown it can do.
A key rival is missing from the benchmarks
Decision-1 uses Qwen3.5-9B as its base. Microsoft said it was the most accurate model the company tested across 36 benchmarks, which together cover nearly 150,000 questions.
Microsoft also said Decision-1 runs 2.5 times faster than H2O-Lightning-4B, which came second. The comparison leaves out Cloudflare’s open-source Clef models, though. Those are also built on Qwen, so they are the closest competitor Microsoft could have tested against. Anyone who wants that head-to-head will have to run it themselves.
How to access it
Decision-1 can be used through Microsoft Foundry and OpenRouter. Because output tokens are free, the cost comes down almost entirely to how much context you send in.
A tough stretch for the startup that started it
Jev started the decision model trend in mid-September, and the concept caught on fast. Bigger companies then took the idea and went further with it, using open small language models such as the Qwen base underneath Microsoft’s model.
In less than a month, Jev has gone from creating a category to competing in it against Microsoft, OpenAI and Cloudflare. On top of that, Microsoft is charging less than five cents per million input tokens.