Why Microsoft Is Turning to DeepSeek V4 for Its New Copilot Cowork Enterprise Agent

Under Microsoft's old pricing model for similar services, billing was based on abstract "premium request units." That model broke under the strain of multi-step agents, where a single user request could trigger a cascade of expensive operations. The switch to consumption-based billing with Copilot Credits—where roughly $0.01 of cost is tied to every interaction's real token usage—directly exposes inference costs to customers . In this transparent pricing environment, any reduction in the underlying model's inference cost directly benefits Microsoft's bottom line, as it can offer lower credit-burn rates while remaining profitable .

Enter DeepSeek V4: Why Microsoft Is Interested

The Axios report that Microsoft is considering a fine-tuned, Azure-hosted version of DeepSeek V4 points to three strategic motivations:

1. Architectural Cost Efficiency: DeepSeek V4 uses an efficient attention mechanism that reduces deployment costs at scale, making it well-suited for high-volume enterprise agent workloads . DeepSeek models are already available in Microsoft's Azure AI Foundry catalog, including V4 Pro and V4 Flash variants, giving Microsoft operational familiarity with them .
2. Supply-Chain Diversification: Copilot Cowork launched with a model picker featuring Anthropic's Claude Opus and Sonnet models, along with OpenAI's GPT-5.5 . By also offering a high-performance, low-cost open-source option as a first-party service, Microsoft reduces its dependency and the pricing leverage held by its frontier model partners.
3. Compliance and Data Sovereignty: Keeping all inference on Azure is a critical compliance advantage. Anthropic's Claude models run on Anthropic's own US-based infrastructure, which has raised questions about GDPR and EU Data Boundary compliance for European enterprises . A Microsoft-hosted DeepSeek V4 would keep all data and processing within Azure's sovereign cloud boundaries, simplifying governance for regulated industries .

How Copilot Credits Pricing Works

With the new consumption model now live, every Copilot Cowork interaction costs money. Here’s the breakdown of the new currency, Copilot Credits:

• Unit Value: One Copilot Credit equals $0.01 of consumption value .
• What It Measures: Usage is calculated at the token level, including input tokens, output tokens, and cached tokens, based on each model's published API rates .
• Payment Options:
  • Pay-as-you-go: Billed at $0.01 per credit with no upfront cost, ideal for low or unpredictable usage .
  • Credit Packs: Pre-purchased in blocks of 25,000 credits for $200 per month, designed for predictable, higher-volume workloads .
• Management: All spending is monitored and controlled through a centralized Cost management dashboard in the Microsoft 365 admin center .

When a lower-cost DeepSeek model eventually appears in Copilot Cowork's model picker, its lower token costs will translate directly into fewer credits consumed for the same enterprise task.

A New Competitive and Security Landscape

This move has ripple effects beyond just Microsoft's accounting department. For enterprise CISOs, it’s a mixed bag. Copilot Cowork already honors existing Microsoft 365 security controls and only accesses data the user is permitted to see, with integration for Microsoft Purview for audit and governance . An Azure-hosted model fits neatly into this architecture. However, the very ability of Cowork to access data across SharePoint, Teams, and Excel simultaneously requires companies to ensure their permissions models are airtight, a new governance challenge in itself .

From a market perspective, adding a high-performance, open-source model as a first-class option in the world's largest enterprise productivity suite is a massive endorsement of the open-weight AI movement. It places direct pricing pressure on competitors like Google Workspace's Gemini and Slack's AI integrations, who will now be benchmarked against Microsoft's ability to offer affordable, complex agentic action at scale. The move signals that the enterprise AI race is now as much about inference cost as it is about raw model capability.