OceanBase tops Berkeley's data-agent benchmark with GLM-5.2
Two benchmark results and one hype check to start the week: a Chinese database outfit took the top of a Berkeley-run data-agent leaderboard on a Chinese model, a frontier model reportedly broke a century-old naval cipher, and the robot industry's own trade body published the least flattering number of the year.
OceanBase, the distributed database business Ant Group spun out, has taken first place on the Data Agent Benchmark with 90.62% accuracy — the first submission to clear 90%, and it was built on a Chinese model, Zhipu's GLM-5.2. DAB is run by UC Berkeley's EPIC Data Lab together with Hasura's PromptQL, and it tests something harder than Text-to-SQL. A single task requires the agent to find the relevant tables across PostgreSQL, MongoDB, SQLite and DuckDB, plan an analysis path, run the joins and calculations, then verify the answer — across finance, biomedicine, intellectual property, government and media datasets. OceanBase's entry, internally codenamed Scout, runs a three-stage loop: profile the data through its DataLens component, execute a planned query path, then check the result against tracked evidence and retry when the check fails.
The company's claim is that the win validates the combination rather than the model: GLM-5.2 handled reasoning while the database handled discovery, correlation and result validation, and the submission outscored several entries built on GPT, Claude Opus and Claude Fable. Read it with the usual discount — this is a vendor reporting its own submission — but the leaderboard itself is not the vendor's, and the interesting claim is architectural, not model-specific. Databases spent a decade being told to hand their data to AI; the pitch here is that the database becomes the thing that makes the AI's answer checkable. We have covered the model half of this stack before — Europe's top-scoring AI model is a compressed GLM-5.2 — and the serving work behind it in GLM's Infra Agent built the stack that serves GLM.
A researcher says GPT-6 Astra deciphered a German naval radio message from 1918 that had resisted codebreakers for 108 years. The message used the ADFGVX cipher, and the researcher's own account of why it survived this long is the interesting part: the key was believed to have entered service only after the message was sent, so nobody tried it. The model reportedly searched historical archives on its own, reconstructed the transposition, and then checked its plaintext against the logs of the cruiser HMS Canterbury to confirm the reading. Turing-adjacent bragging aside, the durable signal is the shape of the work — long-horizon search, cross-referencing scanned documents, and self-verification against an independent record. It is one researcher's claim and has not been reviewed by cryptographers, so treat the decipherment as plausible rather than settled.
The International Federation of Robotics counted roughly 7,000 humanoid robots sold worldwide in 2025 for industrial and professional use — and says a large share went to research rather than productive work. That is the trade body's own tally, and it lands badly against the shipment targets vendors have been putting in their decks: Tesla has been auditing suppliers for tens of thousands of Optimus units next year, and UBTECH is running a line it says can build a humanoid every ten minutes. Both things can be true — a supply chain gets built years before the orders arrive — but the IFR number is the first hard denominator anyone has published for how many of these machines are actually out there doing a job.
What to watch: whether OceanBase publishes Scout's harness so other teams can reproduce the 90.62%, and whether cryptographers confirm the Astra decipherment.
Would you trust a data agent's answer if you couldn't see how it verified itself? Tell us in the comments.
