Snowflake Adds Dynamic Model Routing to Cortex AI Gateway

Enterprise teams running AI agents at scale are discovering that no single AI model is ideal for every task. Advanced models can be unnecessarily expensive for simple questions, while smaller models may struggle with complex requests. AI model routing addresses this challenge by automatically selecting the most suitable model for each task.

Snowflake’s Cortex AI Gateway now supports dynamic model routing, allowing companies to choose an “automatic” option instead of manually assigning a fixed model. The system evaluates each request and routes it to the model that offers the best balance of response quality, speed, and cost.

Snowflake says the technology can reduce token costs by up to three times for certain workloads, based on the company’s internal testing. The company found that simple questions are often sent to highly capable models, resulting in unnecessarily high costs and longer response times.

Snowflake’s announcement comes as the broader AI industry invests in automated model selection. Databricks, AWS, Google Cloud, and Nvidia have all introduced or announced model-routing technologies. However, Snowflake argues that effective routing involves more than comparing model prices and performance. Enterprise AI routing must also account for governance, security, data access, and context.

“Getting context and governance right is critical to building high-quality, enterprise-grade agents,” Baris Gultekin, vice president of AI at Snowflake, told VentureBeat. “Context, trust, and model selection are all closely related.”

Snowflake Uses Two Methods to Route AI Tasks

Dynamic model routing is built into the Cortex AI Gateway, which Snowflake launched in July 2026 as a governance layer for AI agent and model traffic. Before dynamic routing became available, customers selected models from a static list for each task rather than using a true fallback or automated routing system, Gultekin said.

According to Gultekin, Snowflake’s dynamic routing system uses two primary mechanisms:

Small models handle tasks first. Snowflake’s advisor pattern sends a request to a smaller model initially. If that model cannot complete the task, it can call a more capable model as a tool and continue processing the request.

A classifier routes requests based on task history. A classifier trained on historical queries identifies the complexity of incoming requests and directs simple questions to less expensive models while sending more demanding tasks to advanced models.

Customers can restrict model selection. Automatic routing is optional. Organizations can continue to pin requests to a single model or define an approved group of models. The routing system will operate only within those customer-defined boundaries.

Snowflake does not charge an additional routing fee. AI costs are based on token usage. When a request is routed to a lower-cost model, the customer’s bill can be reduced without paying an extra fee for the routing decision.

Enterprise Access Controls Apply to Data, Models, and Agents

Snowflake integrates AI model routing with the same access-control framework used for data governance.

Governance begins at the data layer through role-based access controls. It then extends to models by mapping customer roles to approved model groups. The same approach applies to AI agents, which can be assigned narrower permissions than the users who invoke them.

Open models can run within a customer’s selected region to support data residency requirements. Gultekin said that inference for both open and proprietary models remains inside Snowflake’s security perimeter instead of being sent to an external provider.

This regional and security-boundary control can be especially important for open models developed outside the United States, including DeepSeek-V4-Flash and GLM-5.3, which were both developed in China.

Snowflake’s recent acquisition of Natoma adds another layer to its enterprise AI strategy. The agreement provides managed access to more than 100 Model Context Protocol, or MCP, connectors. This allows agents to use connected tools, such as email, with limited read-only permissions instead of receiving broad access.

Better Context Helps Enterprises Use Less Expensive Models

Snowflake has also announced Horizon Context and Cortex Sense, tools designed to provide AI agents with more relevant information before they process a request.

Without sufficient context, an AI model may need to perform additional exploratory work. For example, it may write and test SQL queries, search for relevant data, or retry a failed operation. This process can be expensive and may require a more capable model to produce a reliable result.

By preparing and packaging the necessary context in advance, Snowflake aims to eliminate much of this exploratory work. As a result, a smaller and less expensive model may be able to complete the same task effectively.

Snowflake also incorporates agent memory into the context provided to models. As an agent handles repeated requests, its memory is updated and reused in future interactions. The agent does not need to solve the same problem from the beginning each time. Instead, relevant prior knowledge becomes part of the context sent to the model.

OpenRouter, Databricks, and Nvidia Target the Same AI Routing Challenge

Snowflake is entering an increasingly competitive AI model-routing market. OpenRouter is one of the best-known options, offering organizations a way to route requests based on model cost and performance. Nvidia announced Switchyard on August 11 as a technology layer for selecting AI models. Databricks also offers Smart Routing through its Unity AI Gateway.

“What’s interesting is where differentiation has shifted,” Sanjeev Mohan, president and founder of SanjMo, told VentureBeat. “Snowflake isn’t really selling routing. It’s selling routing that never leaves the managed data perimeter, with access control, tagging, and cost attribution already attached.”

Mohan added that organizations whose data and compliance programs are already centered on Snowflake may benefit most from routing that keeps information in the appropriate environment while attributing spending to the right teams and cost centers. Companies without a central data platform may prefer neutral gateways that can route requests across a broader range of models.

Mohan sees the market developing across three distinct groups rather than as a single competitive category.

Databricks approaches AI governance through data engineering and machine learning lineage. Its Unity Catalog manages data, models, and pipelines for teams that build and train AI systems.

Snowflake approaches governance through analytics and access control. Its platform focuses on determining who can access specific data and models while tracking usage across business units.

The third group includes neutral gateways such as OpenRouter, LiteLLM, and Portkey, along with hyperscaler platforms such as Azure AI Foundry. These providers typically compete on model variety, flexibility, and reduced vendor lock-in rather than deep integration with a specific governance framework.

Choosing an AI Router Also Means Choosing a Governance Strategy

As AI agents become more widely deployed, model routing is becoming an important part of enterprise AI infrastructure. The central decision is not simply which router is fastest or cheapest. Organizations must determine which governance model best fits their existing data, security requirements, and operating structure.

Manual model selection becomes expensive at scale. Selecting models manually may work when a company operates only a few AI agents. However, the approach becomes difficult to manage as hundreds of agents make routine model calls without automated monitoring for cost, performance, and quality.

Companies should evaluate governance capabilities, not just routing features. The most important question is which governance model aligns with the organization’s existing data assets and provides the cost visibility required to prevent unexpected inference spending.

The best starting point depends on where company data already resides. According to Mohan, Snowflake-centric organizations may gain more from in-platform routing that respects existing access controls and assigns costs to business units than from maximum model variety. Databricks-focused teams may prioritize lineage across model training and deployment, making a gateway built around that lineage more suitable.

Organizations operating across multiple platforms, or those that prioritize model choice and minimal vendor lock-in, may be better served by a neutral AI gateway such as OpenRouter.

“Instead of starting at the router, practitioners should start where the commitment to managed data and platforms already exists, and how much margin is exposed to inference costs,” Mohan said.