Microsoft is looking for a way to keep its Copilot Cowork enterprise assistant reasonably priced. And the solution it is considering is attracting attention. The company is testing whether it could hand off some of the work to the Chinese model DeepSeek. According to Axios, this would be Microsoft's own customized version of the DeepSeek V4 model, hosted by Microsoft itself.
Copilot Cowork is currently powered by models from OpenAI and Anthropic. But they are becoming more expensive. Both companies have also abandoned unlimited plans, which had begun costing them too much. And therein lies the heart of the problem: agentic tools like Copilot Cowork call the model again and again as they work, often hundreds of times for a single task. Productivity rises, and so do the bills.
Charles Lamanna, who oversees Copilot, agents, and the entire platform at Microsoft, told Axios: "We have users who complete hundreds of tasks a week, which is great—they are incredibly productive—but the consequence is that costs can rise very high," he explained. According to him, testing revealed one thing: offering Copilot Cowork with unlimited use simply cannot be profitable.
Microsoft is switching to usage-based pricing
So what will the company do about it? It is changing how customers are billed. Copilot Cowork will now operate on a model where customers pay according to how much computing power they actually use. More work, a higher bill. Less work, a lower one. A cheaper version built on DeepSeek also fits into this model. Microsoft says it plans to launch the lower-cost version within the next few weeks. It will only reveal which model ultimately powers it at that point. DeepSeek may not be the only option under consideration, as the company has not ruled out another open model.
If DeepSeek is deployed, the model will run entirely on Azure, Microsoft's own cloud platform. Customer data would therefore remain within its environment and be subject to the same security standards as the rest of Azure. For companies, it would also be a choice, not a requirement. Microsoft adds that it fine-tuned the model itself and added its own safeguards, including adjustments intended to reduce bias in responses.
The entire move aligns with the direction Microsoft has been discussing for some time. Instead of relying on the same models from OpenAI and Anthropic, it wants to build on multiple sources at once. Satya Nadella elaborated on this in a lengthy post on X titled "A frontier without a functioning ecosystem is not stable." In it, he warned that allowing a few giant companies to dominate the entire industry would harm both the economy and society. Microsoft is evidently trying to loosen the tight grip of its biggest partners, especially following its strained relationship with OpenAI.
Will Washington frown upon it?
Choosing a Chinese model comes with one complication. The Trump administration keeps a tight rein on U.S. companies when it comes to foreign artificial intelligence. It has threatened Chinese companies over suspicions that they trained their models on outputs from American companies. It has considered banning certain models, including DeepSeek. And it recently suspended foreign users' access to Anthropic's latest domestic model after security concerns emerged.
So no matter how many safeguards Microsoft adds, enthusiasm from the White House is unlikely. And convincing other companies to follow the same path will be difficult as long as the threat of government scrutiny hangs over anything that does not fall directly under an American company.
According to investors, Microsoft wants Copilot to remain useful while ensuring that model costs do not eat into its margins. If demand for agents continues to grow, a cheaper model could be exactly what helps the company keep customers satisfied.
Sources: gizmodo.com, finance.yahoo.com and seekingalpha.com
