Snowflake has introduced dynamic model routing in its Cortex AI Gateway, allowing enterprise customers to choose “auto” instead of a fixed model for each AI task. The system automatically selects the model that balances quality and cost, and the company’s internal testing shows this can cut token costs by as much as three times on some workloads. The feature addresses a common problem where simple questions are handled by the most capable - and most expensive - model, driving up response costs and latency.
The capability builds on the Cortex AI Gateway, which Snowflake launched in July 2026 as a governance layer for agent and model traffic. Previously, model selection used a static list per task rather than a true fallback system, according to Baris Gultekin, Snowflake’s vice president of AI. Under the new “advisor pattern,” a small model attempts a task first and calls a larger model as a tool only if it cannot finish the job. A separate classifier, trained on past queries, routes straightforward questions to simpler models. Customers can still pin a task to one model or a defined set of models, and routing only works within that boundary.
Snowflake prices AI purely on token usage, so routing to a cheaper model lowers the bill with no extra fee for the routing decision itself. Governance is tied to the same access controls used for data, starting with role-based access and extending to model buckets and agent privileges. Open models can run from a customer’s own region to meet data residency requirements, and all inference stays inside Snowflake’s security boundary rather than routing to an external provider. That setup matters for open models from non-U.S. origins, including DeepSeek-V4-Flash and GLM-5.3, both developed in China.
The company’s recent acquisition of Natoma adds more than 100 MCP connectors with scoped, governed access, such as read-only access to email. Snowflake also announced Horizon Context and Cortex Sense tools to provide context capabilities. Without good context, a model must do exploratory work itself, writing and testing SQL and retrying failed queries, which is expensive and often requires a more capable model. Packaging context in advance removes that step, so a simpler model can handle the same task. Agent memory is also folded into future queries, so the system does not re-solve the same problem from scratch each time.
Snowflake’s move comes amid a broader industry shift toward automated model routing. Databricks, AWS, Google Cloud and Nvidia have all announced some form of routing technology, and OpenRouter remains a widely known platform for routing based on cost and performance. Nvidia announced Switchyard on August 11 as a technology layer for AI model choice, and Databricks offers Smart Routing for its Unity AI Gateway.
Analyst Sanjeev Mohan of SanjMo said the differentiation has moved past routing speed or cost. He noted that Snowflake is selling routing that stays inside a governed data boundary with access controls and cost attribution already attached, which matters for companies whose data and compliance already center on Snowflake. Mohan framed the market as three camps: Databricks governs from data engineering and ML lineage, Snowflake governs from analytics and access control, and neutral gateways such as OpenRouter, LiteLLM, Portkey and Azure AI Foundry compete on model breadth and avoiding lock-in.
Mohan said practitioners should not start with the router but with where their governed data and platform commitment already live, and how exposed their margins are to inference cost. Manual model selection becomes a cost liability at agent scale, he added, since hundreds of agents making routine model calls with no automated cost check can add up quickly.
