Snowflake added dynamic model routing to the Cortex AI Gateway, enabling enterprise teams to choose “auto” mode instead of locking in a single model for all tasks. When this option is enabled, the platform attempts to identify the model that achieves the best balance between answer quality and cost for each query, rather than always sending simple questions to the most powerful and expensive model.
Snowflake said this mechanism reduced token costs by up to threefold for some workloads, based on internal tests conducted by the company. The result is not a general guarantee for every environment, but it illustrates the problem the feature targets: running a large number of intelligent agents can make manual model selection a cumulative source of cost, particularly when advanced models handle tasks that simpler models could perform.
How does the system determine the appropriate model?
The routing mechanism, according to Baris Gultekin, Snowflake’s vice president of AI, relies on two paths. In the first, a small model is initially used in what the company calls “advisor mode.” If the model cannot complete the task, it invokes a larger model as a tool and continues from there.
The second path uses a classifier trained on a history of previous queries to identify straightforward questions and automatically route them to simpler models. Automatic routing remains optional; customers can lock in a specific model or limit the selection to one model or a defined group of models. Snowflake does not charge a separate fee for the routing decision, because its AI services are priced based on token usage; therefore, selecting a less expensive model reduces the bill.
Routing is tied to governance and context
Cortex AI Gateway, which Snowflake launched in July 2026, provides a governance layer for model and agent traffic. The company says access controls apply not only to data but also to models and agents: roles determine which models are permitted, and an agent can be restricted to narrower permissions than those of the user invoking it.
Open models can also be run in the customer’s region to meet data-residency requirements. According to Gultekin, inference operations, whether for open or proprietary models, remain within Snowflake’s security boundaries instead of being routed to an external provider. This point is particularly relevant when using models with non-American origins, such as DeepSeek-V4-Flash and GLM-5.3, both of which were developed in China.
Snowflake’s acquisition of Natoma also adds more than 100 MCP connectors with controlled and governed access. This makes it possible, for example, to give an agent permission to read email only instead of granting it broader permissions over the connected tool.
Why does this development matter?
Snowflake believes that providing context in advance enables less expensive models to complete tasks that previously required a more powerful model. Without sufficient context, a model may have to explore data, write and test SQL queries, search, and try again. Horizon Context and Cortex Sense tools prepare this context in advance, while agent memory is also incorporated into subsequent queries, reducing the need to solve the same problem from scratch each time.
The move comes amid broader competition that includes Databricks through Smart Routing in Unity AI Gateway, Nvidia through Switchyard, which it announced on August 11, as well as OpenRouter, LiteLLM, Portkey, and cloud service provider gateways such as Azure AI Foundry.
According to Sanjeev Mohan, founder of SanjMo, Snowflake’s differentiator is not routing alone, but keeping routing within governed data boundaries while linking it to access controls, usage tags, and cost allocation to teams. In practice, this approach may be better suited to enterprises whose data and governance center on Snowflake, while multi-platform teams may prefer neutral gateways that provide a broader range of models and reduce dependence on a single provider.