Fabric’s advanced NL2SQL is going default—so teams need to re-baseline their data agents

Microsoft has made advanced NL2SQL and advanced DAX generation generally available for Fabric data agents, and it plans to move the core improvements into the standard runtime between November 2 and November 10. The same rollout will switch the standard runtime to GPT-5.6 Terra, with no opt-in required. That makes this more than an optional feature launch: existing agents can change behavior as the rollout reaches their environment. Microsoft Fabric

From one-shot generation to exploration

Advanced NL2SQL does not commit immediately to one query. Microsoft says the agent can inspect the schema, test joins and look up stored filter values before generating SQL. It exposes three operating patterns: following a relevant approved example with substitutions, asking a clarifying question when the request is under-specified, or exploring data and iteratively refining a query when no example fits. Run steps expose the path and assumptions for debugging. Microsoft Fabric

Advanced DAX applies a similar tool-calling loop to Power BI semantic models and uses instance-value indexing to map a user’s language to values actually stored in columns. Microsoft says the resulting behavior is shared across the Fabric data agent, Power BI and Microsoft 365 Copilot experiences. Microsoft Fabric

The benchmark signal—and its limit

On Microsoft’s internal 2026 evaluation set, advanced NL2SQL raised “smart-answer” accuracy from 78.8% to 90.9%. Median end-to-end latency fell from 28.4 seconds to 26.0 seconds, while P95 fell from 46.8 seconds to 32.9 seconds. Its separate internal NL2DAX evaluation moved accuracy from 89.15% to 91.86%, cut P50 latency from 39.1 seconds to 20.4 seconds and cut P95 from 97.5 seconds to 34.7 seconds. Microsoft Fabric

Those are vendor-run results on undisclosed internal evaluation sets, not an independent benchmark. Microsoft explicitly cautions that results vary by workload, data model and configuration. Teams should treat the numbers as a reason to test, not as a portable accuracy guarantee. Microsoft Fabric

What to test before the default changes

The standard-runtime rollout includes improved example following, filter-value substitution and data exploration for SQL; the tool-calling loop and instance-value indexing for DAX; and the model upgrade. Schema descriptions and Markdown-based topics for SQL sources remain preview-only. Microsoft Fabric

Microsoft recommends rerunning existing ground-truth question sets with the Fabric data agent SDK, comparing the new runtime with the previous baseline, inspecting visible reasoning steps and retuning examples or instructions that were written around older behavior. That is the practical takeaway: a more exploratory agent may improve difficult queries, but it can also interpret established guidance differently. Teams should preserve their current baseline and test accuracy, latency and instruction compatibility as the new runtime lands. Microsoft Fabric