Snowflake has added dynamic model routing to its Cortex AI Gateway, enabling enterprises to automatically route each AI query to the optimal model based on task complexity—reducing token costs by up to 3x on certain workloads.
The feature addresses a growing enterprise problem: running AI agents at scale reveals that a single model handles every task poorly. Either the model is too expensive for simple questions or not capable enough for complex ones.
“Enterprises are overpaying for simple AI queries—sending straightforward tasks to premium models that deliver expensive, slower responses than necessary,” said Baris Gultekin, Vice President of AI at Snowflake. “Model routing solves this by matching each task to the right model.”
Two Mechanisms Power Intelligent Routing
The dynamic routing operates through two primary mechanisms:
Advisor pattern: A smaller model attempts each task first. If it cannot complete the job, it calls a larger model as a tool and continues from there—avoiding unnecessary premium model invocation for simple queries.
Task classifier: A separate classifier, trained on historical query patterns, automatically routes straightforward questions to simpler models based on past performance.
Customers retain full control. Auto-routing is optional, and enterprises can pin specific models or restrict routing to a defined set. There is no separate fee for routing—costs decrease naturally when queries route to cheaper models.
Governance Extends to Model Selection
Snowflake ties routing to its existing data governance framework. Role-based access controls that manage data permissions extend to models, where customer roles map to approved model buckets. This extends further to agents, where an agent can be restricted to narrower privileges than the user invoking it.
All inference—both open and proprietary models—stays within Snowflake’s security boundary rather than routing to external providers. This matters for open models with non-U.S. origins, including DeepSeek-V4-Flash and GLM-5.3, which can run from a customer’s own region to satisfy data residency requirements.
The timing aligns with Snowflake’s acquisition of Natoma, which adds over 100 MCP connectors with scoped, governed access—enabling agents to receive read-only access to connected tools rather than broader permissions.
“Model routing is about more than price and performance—it’s about governance and context,” Gultekin noted. “Getting the context and governance right is crucial for building high-quality enterprise-grade agents.”
Databricks, AWS, Google Cloud, and Nvidia have also announced model routing capabilities, signaling a broader industry shift toward intelligent model selection for cost optimization.