Microsoft launches Decision-1 for fast classification and routing tasks

Microsoft is not building a novel decision model architecture; it's repackaging Qwen3.5-9B with optimizations for the constrained-inference category that OpenAI just commercialized. The real news is that Microsoft chose to compete on this axis at all, signaling a shift from general-purpose LLM dominance toward task-specific inference tiers.

OpenAI's Decisions API launched October 6th, pricing constrained classification tasks at $0.10 per million tokens, roughly one-tenth the cost of general inference. Microsoft's Decision-1 enters the same market four days later, targeting identical use cases (routing, classification) with comparable latency claims. This mirrors the broader fragmentation pattern we've tracked: Sophos using Daybreak for security automation, Musubi's PolicyLM-1.7B for moderation, and Reflection's Beam competing on efficiency rather than scale. The difference here is that Microsoft is not leading the category; it's following OpenAI's pricing and positioning playbook into a market segment that didn't exist a week ago.

If Microsoft bundles Decision-1 into Azure's consumption pricing lower than OpenAI's $0.10/M token rate within 60 days, that confirms a margin-compression strategy to lock enterprise customers into the Azure stack. If the model's 83.5% accuracy benchmark fails to replicate on held-out datasets outside Microsoft's control, that suggests the numbers are tuned to the test set rather than production robustness.