Snowflake adds smart model routing to cut AI costs by up to 75%

AI spending is rising faster than measurable business value, prompting organizations to reconsider whether every task truly needs a state-of-the-art model. Snowflake's response centers on two linked capabilities. Dynamic routing removes the friction of manual model selection by letting administrators set quality thresholds and compliance rules once, then allowing the system to pick the lowest-cost model that meets those requirements for each request. As the routing layer learns from real-world workloads, it can adapt to changing model capabilities and pricing without requiring teams to rebuild applications.

The expansion of open-model access strengthens that routing layer by creating a wider pool of cost-effective options. DeepSeek-V4-Flash 0731's 74.4% score on ADE-bench—matching or beating proprietary models in data engineering tasks—signals that open-source models have reached parity on the benchmarks that matter most to enterprise data teams. GLM-5.3's strength on ADE-bench with minimal token footprint appeals to organizations running high-volume, self-hosted workloads where cumulative inference cost drives total spend.

By operating the full stack—data, compute, model weights, and agent orchestration—Snowflake avoids latency and cost penalties from data transfer and can optimize routing for patterns unique to enterprise workloads rather than generic API endpoints. The combination of automated routing and a growing pool of capable, efficient models creates a feedback loop: as new models arrive and existing ones improve, a larger fraction of requests can be handled by cheaper models while business outcomes remain constant, allowing costs-per-outcome to decline over time.