Hebbia says GPT-6 Astra beats rivals on finance citations and decks

OpenAI has published a customer story about Hebbia, whose product Matrix helps bankers bring financial research together and links answers to the documents behind them. Hebbia tested OpenAI's GPT-6 Astra on two parts of that work: finding the sources a question requires, and building a presentation that follows the banker's brief.
The headline numbers are Hebbia's own. On citing the required types of sources, GPT-6 Astra scored 13.6 percentage points above the best alternative tested, and 99.5% of the excerpts it cited were relevant to the question. On presentation, it met 97.9% of the applicable requirements, 14.4 percentage points ahead of the best alternative tested on that measure. The story says other models traded off polish against adherence to the brief, while GPT-6 Astra ranked first on both.
The testing conditions are described briefly. Hebbia tested a prerelease build in August-September 2026, using the same financial research cases and grading criteria for each model. Domain experts wrote or audited the criteria against source documents.
The story illustrates the use with a hypothetical example. A banker is helping a client assess a strategic acquisition of an industrial-software company, with roughly 2,000 files to review. The banker could load the data room into Matrix and examine customer concentration, contract terms and churn in cited cells. Matrix could bring in peer filings and industry news to help develop the deal thesis, then turn the research into a diligence summary and valuation deck in the firm's template. In that workflow, GPT-6 Astra selects documents or Matrix cells to support an answer, then turns the analysis into native PowerPoint slides that follow the user's structure and guidelines.
Traceability is central to the pitch. Users can open a citation to see the underlying document and highlighted passage, and slide footnotes lead back to the source material. A review could be as specific as checking a customer-concentration figure against the source table and correcting the reporting period before the deck goes to the investment committee.
Kavita, the Hebbia speaker quoted throughout, makes several points. Users base millions of dollars of decisions on what the AI returns, so citations matter for traceability and for confidence. With thousands of documents in a data room, even a single percentage-point gain can yield a lot of marginal benefit. On slides, she does not want to send a partner a deck that looks like AI made it; she wants it in the firm's branding and looking like a human-made slide. She says earlier models were strong on craft or on following guidelines but not both, and that GPT-6 Astra is the first she has seen that is number one on both.
Hebbia wants bankers to spend less time reformatting decks and checking sources, and more time assessing a deal. Kavita expects the hours saved to be significant, saying overnighters get reduced to a few minutes of reviewing the output. She also sees a future role for agents that monitor information in the background and surface signals for finance professionals to assess. Hebbia says it wants to build on these results to help finance professionals assess deals with research they can trace back to its sources.
Key facts
- In Hebbia's tests, GPT-6 Astra scored 13.6 percentage points above the best alternative tested on citing the required types of sources; 99.5% of the excerpts it cited were relevant to the question.
- It met 97.9% of applicable presentation requirements, 14.4 percentage points ahead of the best alternative tested on that measure.
- Hebbia tested a prerelease build in August-September 2026, with the same financial research cases and grading criteria for each model; domain experts wrote or audited the criteria against source documents.
- In the described workflow, the model picks documents or Matrix cells to support an answer, then builds native PowerPoint slides that follow the user's structure and guidelines.
- The banker scenario with roughly 2,000 files is Hebbia's hypothetical example, and the monitoring agents are a possible future role, not a shipped feature.
Why it matters
Finance work lives on two things: whether an answer can be traced to the right documents, and whether the output looks like something a banker could send to a client. The story reports that GPT-6 Astra led on both in Hebbia's tests, where Hebbia says other models traded off polish and adherence to the brief. Kavita puts it as the first model she has seen that is number one on both fronts. The gaps are stated as absolute percentage-point leads over the best alternative tested: 13.6 points on citing required source types and 14.4 points on presentation requirements met.
Who it affects
Bankers and other finance professionals who turn large data rooms, financial statements, customer contracts and management presentations into diligence summaries and valuation decks. It also concerns Hebbia's Matrix users, and teams comparing models for document-heavy research with citations. Hebbia says it wants bankers to spend less time reformatting decks and checking sources, and more time assessing a deal.
How to use it
The story describes a workflow inside Hebbia's Matrix, not a standalone recipe. The model selects documents or Matrix cells to support an answer, then turns the analysis into native PowerPoint slides that follow the user's structure and guidelines, in the firm's template. Users can open a citation to see the underlying document and highlighted passage, and slide footnotes lead back to the source. A human review step is part of the picture, such as checking a customer-concentration figure against the source table and correcting the reporting period before the deck goes to the investment committee. The source gives no pricing, availability date or release date for GPT-6 Astra; the tested model was a prerelease build.
How solid is it
Treat it as vendor-reported. The results are Hebbia's own tests, published by OpenAI on its customer-story page. The conditions are stated in outline: a prerelease build tested in August-September 2026, the same cases and grading criteria for every model, and criteria written or audited by domain experts against source documents. The source does not say whether the results were independently verified. The best alternative model is not named, and its absolute scores are not given, only the percentage-point gaps. The number of test cases, documents or graders is not stated.
Risks and caveats
The comparison is against an unnamed best alternative, so readers cannot tell which models were beaten or by how much in absolute terms. The 2,000-file banker scenario is hypothetical and phrased with 'could'. The time savings are Kavita's expectation ('You can imagine the hours this will save'), and no measured figure for hours saved is given. The monitoring-agents idea is a future possibility with no timeline. Citations are also meant to be checked: the story itself describes users verifying figures and reporting periods against source tables before a deck goes out.
“Traceability is really important to our users because they’re basing millions of dollars of decisions on what the AI actually comes back with”
— Kavita, Hebbia, in OpenAI's customer story