Enterprises running AI agents at scale face a trade-off when relying on a single model: high-end models are often too expensive and slow for simple queries, while cheaper models may lack the capability required for complex tasks. Automated model routing, which selects the right model per task, has become a common solution.
Snowflake has added dynamic model routing to its Cortex AI Gateway so customers can choose an “auto” option instead of a fixed model. The system will route each task to the model that offers the best combination of quality and cost. Snowflake said its internal testing shows this capability can reduce token costs by as much as 3x on some workloads. According to Baris Gultekin, Snowflake’s vice president of AI, model selection previously relied on a static per-task list rather than a true fallback system.
Two mechanisms determine routing
Gultekin described two mechanisms that power dynamic routing:
- A small model tries first: under an “advisor” pattern, a smaller model attempts the task; if it fails to complete it, it calls a larger model as a tool and continues from there.
- A classifier sorts by task history: a separate classifier trained on past queries automatically directs straightforward questions to simpler models.
Auto routing is optional. Customers can still pin a single model or restrict routing to a defined set of models; the system will only route within that boundary.
Pricing and fees
Snowflake charges for AI strictly on token usage. Routing to a cheaper model reduces the bill accordingly, and Snowflake does not levy an additional fee for the routing decision itself.
Access controls follow the task
Routing is tied to the same access controls Snowflake uses for data governance. Governance begins at the data level with role-based access controls, extends to models—where customer roles map to buckets of approved models—and extends to agents, which can be given narrower privileges than the invoking user.
Open models can run from a customer’s own region to meet data residency requirements. Gultekin said that inference, whether open or proprietary, remains inside Snowflake’s security boundary rather than being routed out to an external provider. This regional and perimeter setup is important for non-U.S. open models such as DeepSeek-V4-Flash and GLM-5.3, both developed in China.
Snowflake’s recent acquisition of Natoma adds another layer: the deal brings more than 100 MCP connectors with scoped, governed access. For example, an agent could be granted read-only access to a connected tool like email instead of broader permissions.
Context and memory let cheaper models work
Snowflake also announced Horizon Context and Cortex Sense tools that provide context capabilities. Without good context, a model must perform exploratory work itself—writing and testing SQL, searching data, and retrying failures—which is costly and often requires a more capable model. Packaging context in advance removes that exploratory step, allowing simpler, cheaper models to handle the task.
Snowflake also integrates agent memory into the context: as an agent is used repeatedly, its memory updates and is folded into future queries, so the system does not re-solve the same problem from scratch each time.
Competing approaches in the routing market
Multiple technologies address model routing. OpenRouter is a widely known platform enabling organizations to route based on cost and performance. Nvidia announced Switchyard on August 11, 2026, as a routing layer to help model choice. Databricks offers its own Smart Routing as part of the Unity AI Gateway. AWS, Google Cloud and other vendors have also announced routing-related technologies.
Sanjeev Mohan, Principal and Founder of SanjMo, told VentureBeat that Snowflake’s pitch is not simply routing but routing that never leaves the governed data boundary and carries access controls, tagging, and cost attribution with it. For organizations whose data and compliance already center on Snowflake, in-platform routing that preserves data locality and attributes spend to teams can be especially valuable. Organizations without that center of gravity may prefer a neutral gateway that routes across more models with less friction.
Mohan characterizes the market as three camps: Databricks focuses on governance from a data engineering and ML-lineage perspective (Unity Catalog governs data, models and pipelines); Snowflake centers governance on analytics and access control, governing who can touch which data and attributing usage across business units; and neutral gateways (such as OpenRouter, LiteLLM, Portkey and hyperscaler routers like Azure AI Foundry) compete on model breadth and avoiding lock-in.
What customers should evaluate
Model routing has become a baseline requirement for enterprises. The critical decision is which governance model aligns with a company’s existing data and team organization, rather than which vendor’s router is fastest or cheapest.
Manual model selection becomes a cost liability at agent scale. What worked when a team ran a handful of agents breaks down when hundreds of agents make routine model calls with no automated cost checks.
Mohan’s advice: start with where your governed data and platform commitments already live, and with how exposed your margins are to inference cost. For a Snowflake-centric environment, in-platform routing that respects existing access models and bills back to cost centers often provides more value than raw model breadth; teams centered on Databricks and concerned about lineage may prefer a gateway built around that lineage; multi-platform or model-first teams that want maximum choice with minimal lock-in may favor a neutral gateway.
