Snowflake Adds Smart Model Routing to Cut AI Costs Without Losing Quality

The tension between AI performance and cost has forced companies into an inefficient choice: either overpay for powerful models on every task, or build and maintain custom routing logic to pick cheaper alternatives. Snowflake's framing—"intelligence efficiency"—redefines the question: not which model is best overall, but which model is best suited to each specific task at the lowest appropriate cost. The company's answer has two parts.

First, dynamic model routing in Cortex AI Gateway removes the burden from developers. Rather than engineering teams hardcoding model selection, Snowflake's layer observes each request and chooses the lowest-cost option that can confidently complete it. The system respects administrator-approved models and existing data residency rules, so requests stay within compliance boundaries. As the body notes, routing decisions adapt as new models arrive and relative capabilities shift—without forcing customers to rebuild applications.

Second, expanding access to capable open models directly improves the router's options. DeepSeek-V4-Flash 0731 scored 74.4% on Snowflake's internal ADE-bench test, surpassing the proprietary frontier model tested. GLM-5.2 (the predecessor to GLM-5.3) achieved 66% accuracy with the lowest token footprint of any model benchmarked, making it ideal for high-volume, cost-sensitive workloads. By hosting these models directly rather than proxying third-party APIs, Snowflake ensures inference runs inside enterprise governance and keeps data within existing controls—a meaningful advantage for regulated organizations.