On 9 October Microsoft released a decision model post-trained on Qwen3.5-9B. It classifies, routes and scores, giving a probability for each option. Input costs $0.042 per million tokens and output is free, exactly what TypeSafe announced for its Jev on 15 September. The benchmarks are Microsoft's own.
- 9 October 2026: Microsoft announced Microsoft-Decision-1, a model for fast decisions, available in Microsoft Foundry as a public preview and through OpenRouter. The same day Vercel added it to AI Gateway.
- The model is post-trained on Qwen3.5-9B, and Microsoft promises versions on its own models (MAI) and on OpenAI's soon. It answers yes or no, choice and score questions, with a calibrated probability for each option.
- Price: $0.042 per million input tokens, output is free, the same as TypeSafe announced for Jev on 15 September. According to Microsoft, the model has the highest accuracy in its 36-benchmark comparison and is 35 times quicker than GPT-6 Sol.
Below the first numbers in Microsoft's announcement sits an editor's note: the post was updated to add benchmarks for Jev on accuracy and calibration. Jev belongs to TypeSafe and came out on 15 September. Less than a month later Microsoft measures itself against it, and does so in an addition made after publishing.
The price matches to the digit. Per TypeSafe's announcement of 15 September, Jev charges 4.2 cents per million input tokens and does not bill output; the same price still stands in its documentation as of 12 October. TypeSafe itself wrote back then that it can't prove the price isn't subsidised. Microsoft-Decision-1 charges $0.042, which is the same 4.2 cents, and does not bill output either. Through Vercel's AI Gateway the new model can even be called through a TypeSafe-compatible API.
The base says more than the price. Microsoft does not enter the new class with a model of its own: Microsoft-Decision-1 is Qwen3.5-9B, an open model from Alibaba Cloud's Qwen family, post-trained to decide in a single pass. Cloudflare's Clef-flash, which we wrote about on 1 October, sits on the same Qwen3.5-9B. Its own models (MAI) and OpenAI's are promised "soon", with no date.
Except the numbers do not match even between Microsoft's two posts from the same day. On the Command Line blog, the Xbox Research test with more than 10,000 pieces of feedback comes out competitive in quality with GPT-6 Sol, over 14 times faster and 200 times less expensive. On the Foundry blog the description is the same, more than 10,000 pieces of feedback sorted into themes, but the comparison is with GPT-5, and the speed is 80 to 100 times higher. They may be different measurements. The posts don't say.
Microsoft supplies the test itself. The Foundry blog says: where the model's confidence decides what passes automatically and what goes to a person, validate the calibration on your own data. Run the model on a hundred cases from your own queue where you already know the right answer, and see whether 90% confidence really means nine right out of ten.