Snowflake cuts AI costs with smart model routing, open models

AI costs are climbing faster than business value for many organizations deploying multiple agents and applications. The problem is simple: most companies default to running every request through the most powerful (and most expensive) model available, even when simpler tasks—like generating a status summary—do not need frontier-level reasoning. Snowflake's answer is twofold: intelligently route each request to the smallest capable model, and expand the pool of those capable models so the router has more cost-effective options.

The routing layer itself sits within Snowflake's existing enterprise governance controls, so administrators can approve which models are eligible, enforce data residency rules, and audit every routing decision. Because Snowflake maintains the routing logic centrally, customers do not need to rebuild applications or hardcode model selection; the routing layer adapts as new models ship and pricing changes. The internal benchmarks are concrete: on a data-build (dbt) pipeline, dynamic routing achieved up to 3× token efficiency versus frontier-only; on engineering pull requests, teams cut token use by approximately 25% while holding throughput steady.

The open models Snowflake is adding—particularly DeepSeek-V4-Flash, which scored 74.4% on Snowflake's data engineering benchmark and outperformed the leading proprietary model tested—signal that the open-source frontier has caught up where it matters most for data teams. This matters because a larger pool of high-quality, efficient options gives the routing layer more leverage to lower the average cost per task. The compounding effect emerges over time: as routing decisions improve with real-world usage and new models arrive, the share of workloads handled by efficient models grows, while enterprise governance remains intact.