Microsoft is weighing a major shift in its AI strategy as the cost of running advanced AI agents continues to climb. The company has confirmed that it is evaluating DeepSeek V4, along with other open source alternatives, as a potential backend for Copilot Cowork. The move comes at a critical moment, just as Microsoft has rolled out its new agentic AI capabilities to Microsoft 365 customers and introduced a usage based pricing model.
The decision highlights a growing challenge facing the AI industry. While businesses are eager to adopt AI agents capable of handling complex workflows, the cost of operating those systems at scale is becoming increasingly difficult to justify.
Microsoft Searches for a More Affordable AI Model
Microsoft confirmed that it is testing a fine tuned and self hosted version of DeepSeek V4 as a possible alternative to the premium models currently powering Copilot Cowork. The company is expected to make a decision within the coming weeks.
The reason is straightforward. The economics of enterprise AI are changing rapidly.
Leading frontier models from companies such as Anthropic and OpenAI deliver strong performance, but they also come with significantly higher operating costs. DeepSeek V4, on the other hand, offers a dramatically lower cost structure. According to industry estimates, DeepSeek V4 Pro costs around $0.87 per million tokens compared with approximately $50 per million tokens for some flagship AI models.
That difference becomes especially important when AI agents perform multi step tasks. A single workflow may involve dozens of model calls, document searches, tool integrations, and context retrieval operations before delivering a final result. As usage scales across thousands of employees, operational expenses can increase rapidly.
For Microsoft, reducing those costs could make enterprise AI far more sustainable over the long term.
Why Copilot Cowork Is Changing Microsoft’s Cost Equation
Copilot Cowork represents Microsoft’s latest push into agentic AI. Unlike traditional chatbots, these AI agents can complete tasks across multiple Microsoft 365 applications including Outlook, Teams, Excel, and other productivity tools.
The service became generally available worldwide on June 16. Alongside the launch, Microsoft introduced a new consumption based pricing model through Copilot Credits.
Customers still require a Microsoft 365 Copilot subscription, but AI usage now generates additional charges depending on several factors. These include model selection, context retrieval, runtime duration, and the number of actions performed by the AI agent.
This reflects a broader shift in Microsoft’s AI business strategy.
Earlier this month, GitHub Copilot also moved toward usage based pricing. Together, the changes suggest Microsoft is preparing for a future where AI workloads are measured and billed according to actual consumption rather than fixed monthly fees.
The company appears to recognize that agentic AI workloads are fundamentally different from traditional software services. As AI agents become more capable, they also consume significantly more computing resources.
The Political and Security Questions Around DeepSeek
While DeepSeek offers compelling cost advantages, its Chinese origins introduce additional challenges.
Enterprise customers remain highly sensitive to issues involving data security, regulatory compliance, and geopolitical risk. Government scrutiny surrounding Chinese AI technology has intensified over the past year, making any adoption decision more complicated.
Microsoft has attempted to address those concerns by emphasizing that any DeepSeek deployment would be optional and fully hosted within Azure. Customer data would remain under Microsoft’s existing compliance, security, and data residency frameworks.
However, questions remain.
Even if inference takes place entirely on Microsoft’s infrastructure, some organizations may still evaluate risks associated with where the model was developed and trained. For highly regulated industries, those considerations could influence purchasing decisions just as much as technical performance or cost.
A Multi Model Future for Enterprise AI
DeepSeek may not be Microsoft’s only option.
Reports indicate the company is also evaluating other open source and alternative AI models, including offerings from Meta and Mistral. This points toward a broader multi model strategy rather than dependence on a single provider.
Such an approach would allow Microsoft to balance performance, cost, compliance requirements, and customer preferences across different use cases.
At the same time, reports suggest Microsoft has limited the use of certain Anthropic models within Copilot Cowork due to concerns surrounding data retention policies. If accurate, that development indicates Microsoft’s evolving AI strategy is being shaped by more than pricing alone.
Ultimately, Microsoft’s exploration of DeepSeek V4 reflects a larger reality facing the entire AI industry. Businesses want powerful AI agents, but they also need predictable economics. As enterprise adoption accelerates, the companies that can deliver both performance and affordability are likely to gain a significant advantage.
For Microsoft, the next phase of Copilot may depend as much on cost efficiency as it does on AI capability.
Follow TechBSB For More Updates
