Databricks benchmark with Turing shows Genie topping two frontier agents

Databricks published an enterprise data-agent benchmark built with Turing that pits its Genie assistant against two unnamed frontier agents from other labs. The evaluation uses a fictional UK electricity and gas supplier, Calderwick Energy, whose workspace holds hundreds of tables alongside notebooks, dashboards, saved SQL queries and documents; each task ran five times in a live Databricks workspace and was graded against a rubric.

Genie answered 91% of tasks correctly, against 45% and 58% for the two agents, while using under half the tokens and finishing three to four times faster. Databricks attributes the gap to Genie Ontology, the context layer that pulls business knowledge out of those assets. The benchmark is meant to be shareable, and teams can contact their account team to replicate it.