According to media reports, Microsoft is shifting its AI agent product Copilot Cowork to a usage-based billing model and is exploring the introduction of fine-tuned versions of open-source models like DeepSeek V4 to replace the Anthropic and OpenAI models currently called upon daily, thereby drastically reducing operating costs. Charles Lamanna, Executive Vice President overseeing Copilot, agents, and platform business, stated that testing has shown Copilot Cowork “cannot be provided on an unlimited usage basis” — while some users execute hundreds of tasks per week, resulting in significant efficiency gains, “the downside is that costs can become very high.” Microsoft expects to launch more affordable optional models in the coming weeks and emphasized that all models will be fully hosted on Azure cloud, ensuring customer data does not leave the Microsoft cloud environment.
Cost disparity is the core driver: Anthropic’s recently released Fable 5 model costs 50 times more per token than DeepSeek-V4 Pro, while for most daily office tasks, open-source models are equally capable, with overall costs potentially reduced by over 90%. Data platform Vercel disclosed that DeepSeek’s share of the US enterprise market jumped from 1% in April to 17% in May. In a report last week, Citadel Securities noted that the shift to low-cost models is one of the reasons for the recent decline in the AI spending index, stating that “even the most powerful technology must follow the mundane laws of cost curves and marginal returns.” The latest ranking from the LLM arena Arena also shows that among listed open-source models, Chinese models have taken the lead — Zhipu GLM-5.2 (Max) rose to 10th place, DeepSeek-V4 Pro ranked 14th, and the only overseas open-source models remaining at the bottom are Nvidia’s Nemotron 3 Ultra and Google’s Gemma 4.