Microsoft-Decision-1.1 is generally available for routing and classification at $0.042 per million input tokens, with free output
Microsoft is selling a small model built only to choose between options, and pricing it so it can sit in front of every workflow step.
What happened
Microsoft announced on November 17 that Microsoft-Decision-1.1 is generally available in Microsoft Foundry and through OpenRouter and Vercel. It is designed for routing, classification, prioritization, verification and workflow control. It chooses among actions the developer permits, returns calibrated confidence scores, and is trained to abstain when information is incomplete, so the application can act, flag or escalate based on its own thresholds.
The model is built on Google's Gemma 4 26B A4B mixture of experts model and runs on a custom inference engine. Microsoft says it is more than twice as fast as the first version at the same cost, scores 68 on the Artificial Analysis Decision Index v2.1, and now accepts images such as screenshots. In Foundry, input costs $0.042 per million tokens globally and $0.05 in the EU and APAC data zones, and output tokens are free.
Microsoft says more than 4,000 customers used it in the five weeks since preview, for tasks including support triage, tagging and business process automation. In Copilot it decides whether a request goes to Copilot or Cowork, matching a large model's accuracy 2.2 times faster and 4.4 times cheaper, according to Microsoft. The Decoder places it among seven models Microsoft announced at Ignite, aimed at lowering the cost of agents that run all day.
My take
A lot of the model spend in CRM automations goes on small choices: which pipeline a lead belongs in, whether a ticket is billing or technical, whether an email needs a human. Those steps don't need a frontier model, and they benefit more from a confidence score than from fluent text.
The abstain behavior is the part I'd test first. A router that says it doesn't know lets you send the uncertain cases to a person or a larger model, while the rest flows straight through. Before switching, I'd run a few hundred of a client's real records through it, compare against the labels a person already gave them, and set the escalation threshold from that data. The benchmarks are Microsoft's and Artificial Analysis', so your own data is the test that counts.
More posts
- Open OCR models now read 1,000 pages for under $0.10, and better serving saved more than switching modelsNov 20, 2026
- Gemini 4 Neon costs $3 in and $15 out per million tokens and tops Zapier's AutomationBench at 56.3%, though it uses more output tokens per taskNov 20, 2026
- Gemini API's Tool Retrieval loads tool schemas only when the model asks for them, so agents with hundreds of tools send fewer input tokensNov 20, 2026
