Microsoft has officially started Copilot Cowork for general availability worldwide, and in the same breath the tech giant announced that it is actively exploring the deployment of a fine-tuned version of DeepSeek V4 to operate a lower cost tier of enterprise AI agents.
The development, first reported by Axios on Tuesday, signals a significant strategic shift in the way Microsoft thinks about the economics of always-on AI in the enterprise — and raises new questions about model diversity, cost controls and the role of China-developed AI in sensitive corporate environments.
Rising token costs are the root of the problem
The driving force behind this research is clear: using Copilot Cowork in companies is expensive. Charles LamannaMicrosoft’s executive vice president of Copilot, Agents and Platform, told Axios that some enterprise users are running Hundreds of tasks per week through cowork, which caused costs to skyrocket to levels he described “very high.”
Unlike a traditional chatbot, which processes a single prompt and returns a single response, Cowork works fundamentally differently: it plans multi-step workflows, retrieves contextual information, calls external tools, and iterates in loops until a task is fully completed.
Each of these loops consumes model tokens and computing resources, making the overall cost significantly higher than standard AI interactions.
When it is generally available, Copilot Cowork is primarily based on Anthropic models including Opus 4.8 and Sonnet 4.6with OpenAI’s GPT 5.5 available in a separate Frontier preview tier.
Microsoft has not yet officially chosen DeepSeek V4 as a cheaper engine, but the company confirmed that a lower-cost model option will be announced and made available within weeks. In addition to the DeepSeek option, Microsoft is also developing its own solution “Coworking 1” Branded model – described as a secure, in-house solution designed for everyday tasks at a significantly reduced cost.
Azure hosting as a security shield
Microsoft is very aware that the prospect of a AI model developed in China Sitting in an enterprise productivity suite will turn heads – especially among regulated industries, government contractors and security-conscious organizations.
To address these concerns directly, the company has confirmed that any DeepSeek-based model option would do this hosted entirely on Microsoft’s own Azure infrastructurewith all customer data remaining strictly within the tenant boundary and covered by Microsoft’s existing corporate security, compliance and data residency controls.
The company also confirmed that it has refined the model and implemented additional safeguards, including targeted adjustments Reducing bias.
The approach is well known – it reflects the strategy that Microsoft has pursued Azure AI Foundry to offer third-party models under its own operational and contractual umbrella, essentially laundering third-party risk through Microsoft’s own corporate guarantees. However, whether this will satisfy even the most security-conscious corporate customers remains to be seen.
Usage-based pricing is changing how companies calculate AI
In addition to the DeepSeek news, Microsoft’s decision to move Copilot Cowork from blanket licensing to Usage-based pricing through “Copilot Credits” represents a broader and potentially more consequential shift for the enterprise AI market.
The cost of the new model is determined by a combination of model usage, context retrieval, tool calls, and runtime. The tool is shipped disabled by defaultand administrators get detailed spending controls at the tenant, group and individual user levels.
Microsoft has internally claimed that Cowork is this 30 to 40% cheaper than Claude Cowork from Anthropic in conjunction with a Microsoft 365 connector, although the company admits that costs vary significantly depending on the configuration.
The push toward a cheaper model tier sends an unmistakable message: Even with these savings, the economics of deploying always-on AI agents at the enterprise level remains a real barrier to widespread adoption—and that’s what Microsoft is betting on Variety of modelsrather than model exclusivity, is the way to make the math work for its customers.
3 important insights at a glance:
-
Costs are the core problem: Enterprise users running hundreds of tasks per week through Copilot Cowork are encountering unexpectedly high costs due to the token-intensive nature of multi-stage agent workflows – forcing Microsoft to urgently seek cheaper model alternatives such as DeepSeek V4 and its own proprietary “Cowork 1” model.
-
DeepSeek, but on Microsoft’s terms: To address security concerns associated with a model developed in China, Microsoft plans to host each DeepSeek V4 option entirely on Azure infrastructure, keep customer data within company boundaries, and apply its own tuning and bias protections – effectively the same playbook used at Azure AI Foundry.
-
Usage-based pricing is transforming enterprise AI: The move from blanket licensing to Copilot credits-based consumption pricing is an important structural step and shows that Microsoft sees flexible, multi-model pricing – not a one-size-fits-all subscription – as the key to enabling broad enterprise adoption of always-on AI agents.
Relevant Twitter/X links:
Quick links:
