Snowflake releases AI optimization tool to compare model quality and estimated cost
Cortex AI Function Optimization evaluates candidates on the same data and scoring criteria, then lets customers reuse a selected prompt and model.
Loading page…
Cortex AI Function Optimization evaluates candidates on the same data and scoring criteria, then lets customers reuse a selected prompt and model.
Listen to this story
Snowflake made Cortex AI Function Optimization generally available on November 16, 2026, giving customers a way to test alternative prompts and model configurations for custom AI functions. The tool compares quality and estimated cost using the same dataset and scoring metric, helping teams judge whether a lower-cost model is good enough for a workload. Customers can use a chosen configuration directly or package it as a governed, reusable function; the experiment leaves the original function unchanged.
The AI_FUNCTION_OPTIMIZATION experiment evaluates candidate prompts and models against a consistent dataset and metric.
Customers can compare smaller, less expensive models with larger alternatives to see whether added capability improves quality enough to justify the cost.
A selected configuration can be used in an AI_COMPLETE call or packaged with CREATE AI FUNCTION behind a stable SQL interface.
Snowflake customers can now compare alternative prompts and AI models before choosing a configuration for a reusable function. The company made Cortex AI Function Optimization generally available on November 16, 2026. It evaluates candidates against the same data and scoring criteria, putting quality alongside estimated cost rather than treating either measure alone.
The workflow starts with a custom AI function and searches for alternative implementations. Snowflake says it can generate and evaluate different prompts—the instructions supplied to a model—and model configurations, without customers manually rewriting every version. That makes both the instructions and the choice of model part of the experiment.
The AI_FUNCTION_OPTIMIZATION experiment type evaluates model candidates using a consistent dataset and metric. Customers can compare smaller, lower-cost models with larger, more capable alternatives under the same scoring criteria. The comparison is intended to show whether a cheaper model meets the workload’s quality requirements, or whether a more capable model delivers a material improvement.
The output is a selected prompt and model, not just a comparison to inspect. Snowflake provides two routes for using that configuration after optimization, depending on whether customers want to call it directly or package it as a reusable function:
Optimization does not modify the original AI function. Customers can evaluate alternatives and use a selected result while leaving the starting implementation intact; running the experiment is not itself a replacement of that function.
Loading discussion...
Join the conversation
Explain when extra capability is worth paying for.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.