Microsoft Decision-1 launches as fast Qwen-based decision-scoring model

Decision-1 formalizes Microsoft’s bet that decisions, not paragraphs, will be the real volume driver of agentic AI. By post‑training Alibaba’s Qwen3.5-9B into a pure scoring model, Microsoft is offering developers a cheap way to answer the constant yes or no, route or drop, approve or escalate questions that sit inside complex workflows. At roughly four cents per million input tokens and zero output cost, it undercuts general LLMs and even many open models for classification and gating tasks. This matters because once you break an agent into many small steps, the decision budget, not the generation budget, starts to dominate your bill.

Strategically, Decision-1 also signals a more pragmatic, multi‑vendor stance: Microsoft is happy to base a flagship model on Alibaba’s open weights today and promises future versions rebased on its own MAI models and OpenAI systems. That flexibility lets it chase whatever base model offers the best latency and calibration. In the broader race to AGI, decision models like this make it easier to build deep stacks of automated checks and balances around more powerful generators, which both strengthens safety scaffolding and makes large‑scale autonomous agents economically viable. Competitively, it puts pressure on other hyperscalers and labs to ship their own decision layers or risk ceding control of the agent ‘control plane’ to Microsoft.