Particle.news

Snowflake Adds Dynamic Model Routing to Cut AI Inference Costs

The feature is built into Snowflake CoCo and CoWork so existing agents can use automated model choice without rewriting code.

Overview

  • Snowflake announced Tuesday that Cortex AI Gateway now supports dynamic model routing, which selects the most affordable approved model for each step of an agent task.
  • Routing decisions run inside Snowflake’s secure perimeter and use only administrator-approved models while logging each choice for compliance and auditing.
  • Snowflake expanded its pool of self‑hosted open models, offering DeepSeek‑V4‑Flash 0731 in private preview and saying GLM‑5.3 will be available soon for customers to route to.
  • The company reported internal tests showing token‑efficiency gains—up to about three times on a dbt pipeline and roughly 25% fewer tokens on a coding workload—but those results come from Snowflake’s own benchmarks.
  • By matching lower‑complexity tasks to cheaper models and keeping inference next to governed data, Snowflake expects enterprises to lower per‑task AI costs and speed outcomes without reengineering agent workflows.