Foreign media reports that Microsoft is readjusting its approach to enterprise AI. With the shift in the Copilot Cowork billing model and the inclusion of low-cost open-source models as options, the cost pressures of enterprise AI are coming to the forefront.

Microsoft took the initiative

Microsoft is considering introducing a tweaked version of DeepSeek V4 into Copilot Cowork to replace some OpenAI and Anthropic models. At the same time, Copilot Cowork is shifting from unlimited use to a pay-per-computation model.

Charles Lamanna, a Microsoft executive in charge of Copilot, said that some users complete a large number of tasks each week, but costs also rise rapidly. Based on this, the article concludes that Microsoft is no longer accepting unrestrained model calls.

Token costs become a focal point

The article cites the Sil token issuance and the n Data Token index, stating that the price of the related token has fallen for 12 out of 13 consecutive trading days, nearing recent lows. The author believes this reflects a growing sensitivity among businesses to call costs.

The article also mentions that cutting-edge models are more powerful, but also more expensive. For example, Fable 5's output token cost is said to be significantly higher than Opus 4.8, and the cost gap for similar tasks is widening.

Model routing becomes a necessity

The article argues that what enterprises truly need is not the single strongest model, but the ability to "route models," that is, to select the most suitable and cheapest model based on task complexity.

Arvind Jain of Glean also stated that the bottleneck for enterprise AI lies not only in the model itself, but also in systemic aspects such as retrieval, tool invocation, memory management, and multi-step inference. Microsoft's introduction of a low-cost model is essentially about building a routing mechanism.

Satya Nadella also suggested that companies should build both "token capital" and "human capital" simultaneously. The article concludes by pointing out that whoever can preserve their accumulated knowledge when changing the underlying model will have more control.