Snowflake makes AI function evaluation and optimization generally available
Snowflake has made AI_FUNCTION_EVALUATION and AI_FUNCTION_OPTIMIZATION generally available, giving teams native tools to measure Cortex AI Function quality and compare prompt-and-model alternatives on labeled data.
Snowflake has made AI_FUNCTION_EVALUATION and AI_FUNCTION_OPTIMIZATION generally available, adding native tools for measuring Cortex AI Function quality and searching for prompt-and-model configurations that balance quality with inference cost.
In an evaluation, an experiment runs an AI function or Cortex AI call against a labeled Snowflake dataset using a selected scoring metric. The result combines an aggregate quality score with cost and token usage. Teams can create and run the experiment in SQL or use the guided AI Function Studio interface, whose create, evaluate and optimize workflow became generally available on November 16.
Once a baseline is in place, an optimization experiment generates alternatives by varying the prompt and model configuration. It evaluates every candidate on the specified dataset and metric, recording quality and estimated cost. Snowflake says the feature became generally available on November 16.
The optimizer leaves the original function unchanged. To use a chosen result, a team must materialize it as a new AI function. Evaluation and optimization both use Snowflake EXPERIMENT objects, so users can define a specification and execute it directly in SQL.
The function-level workflow is separate from Snowflake’s Cortex Training preview for managed benchmarking, post-training and inference, which covers model training rather than reusable AI-function experiments.
Snowflake’s synthetic example compared an optimized Haiku candidate, with a validation score of 0.98, against an Opus baseline that scored 0.95. The company calculated 89% lower inference cost for the candidate. This was not a reported customer saving: Snowflake says the scenario and data are synthetic and do not predict or guarantee results or cost savings.
The functions were previously available in public preview, announced in September. Snowflake’s documentation says evaluation and optimization are natively available in select regions and available to accounts in all regions through cross-region inference. Model authorization, regional availability and pricing can still limit the configurations an account can test.
More news

Snowflake adds generally available TwelveLabs video AI to Cortex functions

Snowflake adds Gemini 4 Argon to Cortex AI in public preview

Snowflake brings Claude Sonnet 5.5 to Cortex AI in public preview
