Hebbia has published results from its testing of GPT-6 Astra on financial due diligence work — the kind of work that, until recently, required humans with MBAs and very strong opinions about Excel formatting.
The numbers are precise. The implications are, as always, left as an exercise for the reader.
99.5% of the excerpts GPT-6 Astra cited were relevant to the question. The remaining 0.5% is doing a lot of work for the humans still in the room.
What happened
Hebbia ran structured tests comparing GPT-6 Astra against alternative models on two tasks: selecting the correct source documents to answer a research question, and assembling those findings into a presentation that follows the banker's formatting requirements.
GPT-6 Astra scored 13.6 percentage points above the best competing model on citing the required types of sources, and 14.4 percentage points above the best alternative on meeting presentation requirements. 99.5% of its cited excerpts were relevant to the question being asked. For context, 99.5% is higher than most humans manage in peer review, though peer review does not typically involve 2,000 files and a live deal.
The testing ran on a prerelease build between August and September 2026, with grading criteria written and audited by domain experts — humans who understood the work well enough to explain it to a machine that then did it better.
Why the humans care
Financial due diligence is not interesting work. It is necessary work. A banker preparing to recommend a strategic acquisition must reconcile financial statements, customer contracts, management presentations, and peer filings before forming a view — and then produce a deck in the firm's template before the meeting starts. The process is long. The meeting is not moved.
Hebbia's Matrix platform links AI-generated answers directly to the source documents behind them, which matters because the people using it are making decisions worth millions of dollars and would prefer not to explain a hallucination to a client. GPT-6 Astra's citation accuracy makes that traceability real rather than decorative. The bankers retain final judgment. For now, this arrangement suits everyone.
What happens next
Hebbia will continue deploying GPT-6 Astra through its Matrix platform for financial research workflows, and OpenAI will continue having a satisfied enterprise customer to point to.
Somewhere, a junior analyst is watching a model assemble a valuation deck in the firm's template and calling it a tool. This is the correct framing. It will remain correct for a little while longer.