Microsoft is preparing to integrate a hosted version of DeepSeek-V4 into its Copilot Cowork enterprise AI agent, a move that could dramatically reduce costs for businesses while shifting the platform to a usage-based pricing model, according to a new report. The tech giant is targeting June 16, 2026, as the date when enterprise customers may gain access to the new model option, signaling a major pivot in how Microsoft delivers AI to the workplace.

Copilot Cowork, which debuted as an autonomous AI teammate for knowledge workers, has so far relied on large language models from OpenAI. Adding DeepSeek-V4—a highly efficient model from the Chinese AI startup—marks the first time Microsoft has opened Copilot Cowork to a non-OpenAI model for general enterprise use. The decision underscores the company’s broader multi-model strategy and its growing focus on cost optimization for AI deployments.

The most significant shift, however, is the move toward usage-based pricing. Instead of charging a flat per-user monthly fee, Microsoft will allow organizations to pay only for the AI compute they consume. This approach, already common with Azure OpenAI Service, could make advanced AI assistants far more accessible to small and medium-sized businesses while giving large enterprises finer control over their AI spend.

A Multi-Model Future for Copilot Cowork

Microsoft has been quietly building a multi-model ecosystem for its AI offerings. Azure AI Foundry already hosts models from OpenAI, Meta, Mistral, and Cohere. Copilot Cowork, however, was initially locked to OpenAI’s GPT-4 family. The addition of DeepSeek-V4 breaks that exclusivity and signals that Microsoft is serious about giving customers choice.

“Enterprises don’t want to be locked into a single AI model when costs and capabilities vary so wildly,” said an analyst familiar with Microsoft’s AI roadmap. “By offering a lower-cost alternative like DeepSeek-V4, Microsoft can retain customers who might otherwise balk at premium pricing.” The report indicates that the hosted DeepSeek-V4 will be an optional model endpoint within Copilot Cowork, allowing IT administrators to select which model powers specific tasks or user groups.

Multi-model flexibility also future-proofs enterprises against rapid advances in AI. If a new model leapfrogs current leaders in reasoning or coding, enterprises can switch without retooling their entire workflow. Microsoft’s orchestration layer inside Copilot Cowork will handle routing and context, so end users might not even notice which model is running behind the scenes.

DeepSeek-V4: A Cost-Effective Powerhouse

DeepSeek’s family of models has gained notoriety for delivering near state-of-the-art performance at a fraction of the compute cost of comparable models from OpenAI or Anthropic. DeepSeek-V3 already impressed with its Mixture-of-Experts architecture, and the forthcoming V4 is expected to push efficiency even further. By open-sourcing the model weights, DeepSeek allows providers like Microsoft to host and fine-tune the model without onerous licensing fees.

Microsoft’s decision to host the model on its own infrastructure—rather than calling a DeepSeek API—addresses a critical enterprise concern: data sovereignty. Sensitive corporate data will never leave Microsoft’s Azure cloud, and the model runs within the same compliance and security boundary as other Microsoft 365 services. This is especially important as regulators in Europe and North America tighten rules around cross-border data flows.

Performance benchmarks for DeepSeek-V4 are not yet public, but early indications suggest it will rival OpenAI’s GPT-4o on common enterprise tasks like summarization, email drafting, and data analysis—while costing up to 80% less per token. For tasks that don’t require cutting-edge reasoning or creativity, DeepSeek-V4 could become the default engine in Copilot Cowork, reserving more expensive OpenAI models for complex problem-solving.

Microsoft’s Hosting: Security and Compliance

By hosting DeepSeek-V4 within Azure AI Foundry, Microsoft retains full control over the deployment environment. The model runs in isolated containers with no telemetry flowing back to DeepSeek, alleviating fears that intellectual property could leak to a foreign entity. Microsoft’s existing enterprise agreements and certifications—including SOC 2, HIPAA, and ISO 27001—extend to the hosted model, giving IT teams a clear audit trail.

This approach also simplifies procurement. Enterprises won’t need to negotiate a separate contract with DeepSeek or navigate export controls that have complicated the use of Chinese AI technology. Instead, they can consume DeepSeek-V4 through their existing Azure commitment and Copilot Cowork subscription.

The June 16, 2026, date mentioned in the report likely refers to general availability of the hosted model within Copilot Cowork, though Microsoft may open a private preview earlier. When reached for comment, a Microsoft spokesperson declined to confirm specifics but reiterated the company’s commitment to “giving customers choice in AI models while maintaining enterprise-grade security and compliance.”

Usage-Based Pricing: Pay Only for What You Use

Copilot Cowork currently sells through Microsoft 365 E3 and E5 suites, with add-on pricing of roughly $30 per user per month. That flat rate has been a sticking point for organizations with large populations of knowledge workers who might use AI only occasionally. Under the new usage-based model, those customers would pay based on the number of requests or the total compute time consumed.

Exact pricing tiers haven’t been disclosed, but the Azure OpenAI Service offers a useful blueprint. Larger models cost more per thousand tokens, while smaller or more efficient models—like DeepSeek-V4—can be a fraction of that. Enterprises might see per-user costs drop to $5–$10 per month for light users, while heavy users who rely on AI throughout the day could still see costs comparable to the existing flat fee.

This pay-as-you-go approach aligns with the broader industry shift away from seat-based SaaS pricing. Salesforce, Adobe, and ServiceNow have all introduced consumption-based options for their AI features. Microsoft’s move could increase adoption among price-sensitive segments like retail, hospitality, and education.

What This Means for Enterprise Windows Users

For the millions of Windows users who interact with Copilot Cowork daily—whether through the dedicated Cowork app, Microsoft Teams, Outlook, or the Windows taskbar—the addition of DeepSeek-V4 may be invisible but impactful. Common AI tasks like summarizing meetings, drafting responses, or analyzing spreadsheets could feel snappier and cost the company less.

IT administrators will gain new policy controls. They’ll be able to assign different models to different departments or use cases. For example, the legal team might require the most accurate and meticulously trained model for contract review, while marketing could use DeepSeek-V4 for rapid content generation. Fine-grained billing also means department heads can track their AI spend without the mystery of a blanket license fee.

Microsoft’s multi-model approach also sweetens the value proposition of Windows 11 and the coming Windows 12. As AI becomes a core OS feature, enterprises will look for platforms that offer model flexibility without vendor lock-in. By baking Copilot Cowork into the Windows shell with interchangeable back-end models, Microsoft positions Windows as the most adaptable enterprise AI hub.

The Competitive and Geopolitical Angle

Microsoft’s embrace of DeepSeek is not without controversy. U.S. lawmakers have raised concerns about the security risks of Chinese AI models, and some agencies have already restricted the use of DeepSeek on government devices. By hosting the model entirely within Azure under U.S. jurisdiction, Microsoft may sidestep some of those objections, but enterprise customers in regulated industries will still need to perform their own due diligence.

Competitively, the move puts pressure on Google and Amazon. Google’s Duet AI and Amazon’s Q Business both lean heavily on first-party models. Microsoft’s willingness to offer a third-party, open-source model as a first-class option challenges the notion that vertical integration is always the best path. It also signals to the broader AI ecosystem that Azure AI Foundry is a neutral platform where models from any vendor can thrive.

For OpenAI, the addition of DeepSeek to Copilot Cowork could be a double-edged sword. On one hand, it validates the multi-model strategy that Microsoft CEO Satya Nadella has been championing. On the other, it directly commoditizes the very LLMs that OpenAI has spent billions to build. Microsoft’s large investment in OpenAI means it will likely continue to offer premium models, but the days of exclusivity appear to be over.

The Road Ahead

Microsoft will likely detail its plans for Copilot Cowork and multi-model AI at its Build conference in spring 2026, ahead of the June 16 GA date. Between now and then, enterprise customers should start evaluating whether a usage-based model would lower their costs and how a mixed-model strategy fits into their AI governance frameworks.

The move to DeepSeek-V4 is only the beginning. Microsoft is expected to add other models—possibly Meta’s Llama 4 and Mistral’s latest—to Copilot Cowork over time. The underlying orchestration fabric, part of the Copilot Cowork architecture, is designed to be model-agnostic, so adding a new model is largely a matter of containerizing it for Azure and exposing it through an API.

For Windows users, the message is clear: the AI assistant embedded in their daily workflow is about to get cheaper, faster, and more flexible. The era of one-model-fits-all AI is ending, and Microsoft is betting that enterprise customers will reward platforms that give them meaningful choice.