Microsoft Launched Decision-1, Its Decision Model on Qwen: A New AI Category Is Born | inicIA
Direct answer
On October 9, 2026, Microsoft announced Microsoft-Decision-1, a scoring model: given a question with fixed options, it returns calibrated probabilities for each option (e.g., 80% yes, 20% no). It's post-trained on Alibaba's Qwen3.5-9B (open model), costs $0.042 per million input tokens (output free), and claims 85-millisecond latency. It's Microsoft's entry into a category that exploded in weeks — decision models from OpenAI, Perplexity, and Cloudflare — and it solves the most expensive problem of autonomous agents: deciding fast, cheaply, and with an honest number before acting.
What a decision model is (and how it differs)
- Probabilities instead of text. You give it a state and yes/no, multiple-choice, or rubric-scored questions; it answers with probabilities, no free text or written rationale.
- Named use cases from Microsoft: routing, classification, prioritization, verification, and workflow control — including grading AI responses and agent actions against rubrics.
- Available in Microsoft Foundry (public preview) and OpenRouter; hosted API only, no open weights.
- The price says it all: $0.042 per million input tokens. For comparison: OpenAI's Decisions API (launched the same week, on GPT-6 Luna) costs $0.10, and Perplexity $0.02. The category competes in cents because the volume is enormous.
- The claims are the vendor's: 83.5% average accuracy across 36 benchmarks (~147,000 questions) and "35x faster than GPT-6 Sol" — vendor-reported figures; an independent bench ranked it 6th of 14 on 82 questions. Treat it as an initial signal and validate with your own tests.
The strategic detail few noticed
Microsoft — OpenAI's largest investor — built its decision model on Alibaba's Qwen, announcing it will later rebase onto its own models (MAI) and OpenAI. The lesson is architectural, not geopolitical: not even Microsoft marries a single provider. Design your flows to swap models without rewriting everything, because prices and capabilities move every quarter.
Why your business needs a "traffic light" like this
Every agent that acts needs a cheap guardian before the action: does this lead qualify? (yes, 82%), is this reply safe? (no, 91%), does this case escalate to a human? A decision model is that traffic light: fractions of a cent per query, millisecond responses. Without it, you either pay a large model for every micro-decision or let the agent decide blind.
Related links
- What to do when your AI agent makes a mistake
- The 7 types of AI agents for business (and which one yours needs)
- How to choose an AI provider: 7 questions before signing
Frequently asked questions
What did Microsoft announce on October 9, 2026?
Microsoft-Decision-1, a decision-scoring model post-trained on Alibaba's Qwen3.5-9B. It returns calibrated probabilities per option instead of text, for routing, classification, and agent control.
How much does it cost and where is it available?
$0.042 per million input tokens, output free. Available in Microsoft Foundry (preview) and OpenRouter, hosted API only.
Why does it matter for my business?
Because autonomous agents need a cheap, fast guardian before acting — and because it proves the winning architecture is multi-provider: design to swap models without rewriting your flows.
Take the next step
Your agents need a traffic light before acting: fast, cheap, auditable decisions. That's how we design inicIA's flows. Talk to us on WhatsApp: https://wa.me/17866120930
By inicIA
