[Digital Today reporter Chi-gyu Hwang] Snowflake, a global AI data cloud company, said on Aug. 19 it added dynamic model routing to its Cortex AI Gateway.

The company said the routing feature allows enterprise customers to select an "automatic" option instead of locking in a specific model. The system automatically chooses an optimal model for each task by considering quality and cost together.

Snowflake said there were many cases in which high-performance models were used even for simple questions, unnecessarily increasing response costs and slowing speed. It said internal tests showed the dynamic model routing feature could cut token costs by as much as threefold for some workloads.

Databricks, AWS, Google Cloud and Nvidia have also introduced model routing technologies. Snowflake said model routing is directly tied not only to price and performance but also to governance and context.

Baris Gultekin (바리스 굴테킨), vice president of Snowflake's AI division, said, "To build high-quality enterprise-grade agents, you need to properly establish context and governance." He added, "Context, trust and model selection are interconnected."

Snowflake's dynamic routing works in 2 ways. First is the "advisor pattern," in which a smaller model tries a task first and, if it cannot handle it, calls a larger model like a tool to take over.

It also offers a method in which a separate classifier trained on past query histories automatically assigns simple questions to a smaller model. Customers can also limit the routing scope to a specific model or set of models. No additional fee is charged for routing itself, and charges are set only based on token usage.