OceanBase Tops International Data Agent Benchmark With 90.62% Accuracy

OceanBase's Data Agent submission has topped the Data Agent Benchmark (DAB), an international benchmark for data agents, with 90.62% accuracy, making it the first entry on the leaderboard to break the 90% mark, according to a report from QbitAI.

What makes the result notable is the stack behind it. The scheme was built by the domestic database OceanBase on top of the domestic foundation model GLM-5.2, and it finished ahead of several Data Agent systems built on overseas models including GPT, Claude Opus and Claude Fable. The submission carried the internal codename Scout, and the report says its capabilities will be folded into OceanBase DataPilot.

DAB was introduced by UC Berkeley's EPIC Data Lab together with Hasura PromptQL. Its tasks span internet and local services, finance and equities, biomedicine, intellectual property, enterprise operations, government and public administration, and media and entertainment, and they run against PostgreSQL, MongoDB, SQLite and DuckDB.

Classic Text-to-SQL benchmarks mainly ask whether a model can turn a natural-language question into a query. DAB asks something harder: whether an agent can find the right answer inside complex, scattered, differently shaped real data. A single task requires the agent to understand the data, choose what to use, plan an analysis path, run the query and computation, and then verify its own result.

That is why the benchmark measures more than a model. It tests the combined capability of model, agent and data system: the model understands and reasons, the agent plans and executes, and the data system has to support data discovery, joins and computation as well as validation of the outcome.

OceanBase's entry follows exactly that chain. DataLens first builds a data profile and identifies field-level and data relationships; the system then plans an execution path according to task complexity and carries out selection, filtering, joining and computation. Once a result is produced, evidence tracing and answer validation re-examine the computation, and any problem triggers a revised plan and a fresh check, closing a loop of understanding, execution and repair.

For the company, the more interesting shift is what the result says about the database's role. Databases used to store, query and process data; now that AI agents have become a new kind of data consumer, they also have to help AI understand data, connect it and validate analytical results, moving from handing data to AI to helping AI put data to work.

OceanBase describes the DAB result as a validation of its move from database vendor toward AI data platform, with DataPilot as the product direction that will absorb the Scout work. The item was published by QbitAI as an authorised repost of material supplied by Ant, so the finer benchmark details, third-party reproduction and the product timeline are still to be confirmed.

What to watch next: whether independent runs reproduce the 90.62% figure against the same DAB tasks, and how soon the profiling, planning and verification machinery behind Scout actually ships inside DataPilot for enterprise data work.