Nov 18, 2026: Prompt caching for Cortex AI Functions

Prompt caching for Snowflake Cortex AI Functions is now generally available for AI_COMPLETE. Many AI workloads send the same leading content with every request, such as a long system prompt, classification rules, a schema description, or a set of few-shot examples, followed by a short piece of row-specific data. Without caching, that shared content is reprocessed and billed in full on every row.

With prompt caching, Snowflake processes the shared prefix once and reuses it, so you pay a discounted cached input rate for the repeated tokens. Discounts are up to 97.5% off the standard input rate, and long prompts return faster. Prompt caching is automatic and requires no SQL change or manual enrollment.

For more information, see Caching for Cortex AI Functions.