Microsoft evaluates DeepSeek V3 for Copilot to cut agent costs, will offer cheaper tier within weeks

Microsoft is exploring a self-hosted, fine-tuned version of DeepSeek V3 or another open-source model to power Copilot Cowork, the agentic assistant in its Microsoft 365 suite, according to statements to Axios. The company expects to offer a lower-cost model tier within weeks.

The move comes as Microsoft simultaneously shifts Copilot Cowork to usage-based pricing, charging companies for actual compute consumption rather than flat subscription fees.

Agent economics force the shift

Agentic AI tools like Copilot Cowork call language models repeatedly as they work through tasks, creating costs that scale rapidly with usage.

"We have users who do hundreds of tasks a week, which is great, they're way productive, but the consequence is the costs can go very high," said Charles Lamanna, Microsoft's executive vice president for Copilot, agents and platform.

Copilot Cowork currently runs on Anthropic and OpenAI models. Both providers have raised prices and moved away from unlimited usage plans. Microsoft previously metered GitHub Copilot for similar cost reasons.

DeepSeek offers inference cost advantage

DeepSeek V3, released in December 2024, offers significantly lower inference costs compared to frontier models while maintaining competitive performance. The model is open-source and popular with developers seeking cost-effective options.

Microsoft says any DeepSeek deployment would be optional for customers and fully hosted on Azure, maintaining data inside Microsoft's cloud infrastructure under its security and compliance controls. The company claims it has fine-tuned the model and added safeguards, including bias reduction measures.

Political complications

The timing presents political challenges. Washington has discussed banning DeepSeek, sanctioned Chinese AI firms, and recently forced Anthropic to restrict its top models for non-US users in a dispute that required Commerce Department intervention.

Multi-model strategy emerges

The evaluation signals Microsoft's broader shift toward a multi-model approach, reducing dependence on any single AI lab. This marks a strategic change from its exclusive, often tense relationship with OpenAI.

Microsoft emphasized this remains an evaluation, not a final decision. The company will confirm its model selection when the cheaper tier launches.

What this means

The economics of agentic AI are forcing even Microsoft to reconsider its infrastructure choices. When a company with deep pockets and close ties to both OpenAI and Anthropic considers Chinese open-source models for cost reasons, it reveals how unsustainable current agent pricing has become at scale. Microsoft's willingness to publicly name DeepSeek as a candidate, despite the political environment, underscores the financial pressure. The shift to usage-based pricing and cheaper model options will likely accelerate across the industry as agents move from demos to production workloads.