Analysing thousands of files for an M&A deal is a routine headache for bankers. Hebbia’s Matrix platform tries to make this less painful by tying research answers directly to their sources and automating the creation of client-ready decks. In recent prerelease tests, Hebbia put OpenAI’s GPT-6 Astra to work on these tasks, focusing on two metrics: how well the model finds and cites relevant sources, and how closely its presentations match the strict requirements of investment banking deliverables.
The tests were run on real financial research cases, using criteria written and audited by domain experts. Hebbia compared GPT-6 Astra with other leading models, tracking not just accuracy but how well outputs could be traced back to source documents—crucial in a workflow where millions of dollars ride on the details.
Citation quality and source coverage
One of the biggest challenges in financial due diligence is surfacing the exact evidence a banker needs from a huge data room. Hebbia’s workflow involves loading thousands of files—contracts, financials, news, peer filings—into Matrix. The AI then selects supporting documents or spreadsheet cells to answer questions like customer concentration or revenue durability. In these tests, GPT-6 Astra achieved a 13.6 percentage-point lead over the best alternative tested on citing the required types of sources, and 99.5% of the excerpts it cited were relevant to the question.
Traceability was central: every citation links directly back to the underlying material, so users can check figures or correct details before a deck reaches decision-makers. Even a small improvement here, Hebbia notes, can mean less time spent double-checking and more confidence in the AI’s findings.
Presentation adherence and output polish
Banking presentations have to look and feel like they were made by humans, not AI. GPT-6 Astra was evaluated on how well it followed user briefs and matched firm branding. In Hebbia’s results, the model met 97.9% of the applicable presentation requirements, a . Previous models often forced a trade-off between polish and instruction-following, but GPT-6 Astra led on both, making it the first model Hebbia has seen to do so.




