Share with your CIO
Microsoft is weighing DeepSeek V4 as a model option inside Copilot CoWork, its enterprise AI agent platform that launched globally on June 16. The driver is cost: DeepSeek V4 Pro runs at one-third the input cost and one-seventh the output cost of Claude 4.8, which Copilot CoWork currently uses alongside GPT-5.5. Microsoft says customer data would stay on Azure and fall under its own security policies regardless of which model processes the request. A lower-cost model option is expected within weeks.
What this means for your business
Usage-based pricing on frontier models creates a compounding exposure that most enterprises haven’t fully modeled. When employees run dozens of agentic tasks daily through a platform like Copilot CoWork, token costs on Claude 4.8 or GPT-5.5 don’t stay in the pilot budget. Microsoft adding DeepSeek as a cheaper routing option is a direct response to that arithmetic, and it gives your procurement team a real lever when negotiating enterprise AI contracts.
The deeper move here is what Satya Nadella named explicitly: Microsoft does not want enterprises ceding value to a small number of models. That’s not customer-friendliness talking, that’s margin defense. By building a multi-model platform where models compete on price and capability, Microsoft locks in the orchestration layer while commoditizing the models underneath it. Azure becomes the durable asset. OpenAI and Anthropic become interchangeable tenants. CIOs who recognize this structure can use it: push vendors toward model-agnostic contracts and insist on the right to swap underlying models without renegotiating licenses.
The security framing deserves scrutiny. “Data stays on Azure” addresses data residency, but it doesn’t fully answer questions about model weights trained on Chinese infrastructure or inference-time telemetry. The signal worth watching: whether Microsoft publishes a specific DeepSeek security disclosure comparable to its existing sovereign cloud documentation, or whether it leaves enterprise security teams to fill that gap themselves.
Concept deep-dive: Model routing
Model routing is the practice of automatically directing a user’s request to different AI models based on task complexity, cost, or latency requirements, rather than sending everything to one model. It exists because frontier models are expensive and often overkill for simple tasks. Think of it like airline fare classes: the same destination, different prices for different service levels. In Copilot CoWork, routing means a simple scheduling prompt might hit DeepSeek while a complex contract analysis goes to GPT-5.5. CIOs should ask vendors how routing decisions are logged and who controls the routing rules.
Based on reporting from Microsoft Weighs DeepSeek for Copilot CoWork to Cut Enterprise AI Costs, originally published 2026-06-17 05:46:00.
