Particle.news

Snowflake Adds Dynamic Model Routing to Cortex AI and Expands Self-Hosted Open Models

Snowflake says routing will cut AI costs by sending tasks to administrator‑approved models with inference kept inside its governance boundary.

Overview

  • Snowflake introduced dynamic model routing in Cortex AI Gateway to automatically select the most cost‑effective, administrator‑approved model for each task.
  • The company said routing runs inside existing governance controls, respects data residency, logs every routing decision for auditability, and only considers models approved by administrators.
  • Snowflake is expanding its self‑hosted open‑model pool with DeepSeek‑V4‑Flash 0731 in private preview and GLM‑5.3 announced for private preview availability that may change based on model readiness.
  • Snowflake published internal tests showing routing plus certain open models matched quality while using fewer tokens — up to three times better token efficiency on a dbt pipeline test and about 25% fewer tokens on a coding workload — and noted those results are from company testing.
  • The routing feature is integrated with Snowflake CoCo and CoWork so enterprises can apply model choice across agents and workflows, a shift Snowflake says can lower per‑task AI costs and make large‑scale AI use more auditable and easier to manage.