Snowflake's Cortex AI Gateway now routes AI tasks automatically to the cheapest capable model, with internal testing showing up to 3x token cost reduction. The system uses two mechanisms: an advisor pattern where a small model attempts the task first and calls a larger model only if needed, and a classifier trained on past queries that sorts straightforward requests to simpler models. Auto routing is optional, carries no separate fee, and customers can restrict routing to a defined set of models.
The more interesting story is in the governance layer, not the cost savings. Access controls follow the task through three levels: data, models, and agents, with an agent able to hold narrower privileges than the user invoking it. All inference, including open models like DeepSeek-V4-Flash and GLM-5.3, stays inside Snowflake's security perimeter. The Natoma acquisition adds over 100 MCP connectors with scoped permissions. Snowflake also pre-packages context via Horizon Context and Cortex Sense, which removes exploratory SQL work and lets cheaper models handle queries that would otherwise require a more capable one.
Model routing is now a crowded space. Nvidia announced Switchyard on August 11. Databricks has Smart Routing inside Unity AI Gateway. OpenRouter, LiteLLM, and Portkey serve teams that want model breadth without lock-in. Analyst Sanjeev Mohan frames the market as three distinct camps: Databricks for ML lineage, Snowflake for analytics governance and cost attribution, and neutral gateways for multi-platform teams. The full article is worth reading for Mohan's breakdown of how to match a routing decision to an existing data estate, which is the actual decision enterprises need to make.
[READ ORIGINAL →]