Snowflake cuts AI costs with smart model routing and open models

AI budgets are rising but business returns are not always keeping pace—a dynamic Snowflake frames as a mismatch between model capability and task need. Most organizations route every request to the most powerful (and most expensive) model available, even for straightforward tasks like generating summaries. Snowflake's response targets this waste through two levers: intelligent routing and a broader roster of open models.

Dynamic model routing sits inside Cortex AI Gateway and respects the governance rules already in place—role-based access control, data residency, audit logging—ensuring compliance and visibility. The early results are concrete: a data pipeline workload achieved three times higher token efficiency compared to always using a frontier model, and a coding workload maintained throughput while cutting token use by approximately 25%. Critically, the routing adapts as new models enter the market and existing ones improve, so organizations do not need to rewrite applications to benefit from those shifts.

The addition of capable open models—particularly DeepSeek-V4-Flash, which outperformed Snowflake's tested proprietary baseline on data-engineering tasks, and GLM-5.2, which delivered the lowest token footprint in the benchmark—expands the practical options available to the router. The system compounds: more model choices give routing more cost-effective branches to select, and as real-world usage data accumulates, routing decisions become more precise. Over time, a greater share of workloads can shift to efficient models without compromising outcomes.