Hebbia announces that 99.5% of the extracts cited by GPT-6 Astra were relevant during its tests carried out in August-September 2026 on a preliminary version of the OpenAI model. The company evaluated the search for financial sources and the production of presentations in accordance with the guidelines. The time savings mentioned by Hebbia remain prospects.
Citations provide access to supporting documents
In Hebbia, a citation provides access to the supporting document, with the relevant passage highlighted. The slides also include documentary references at the bottom of the page. The control could, for example, relate to customer concentration data compared to the original table, then to the correction of the reporting period before transmission to the investment committee.
The company wants to reduce document formatting and verification tasks to allow more time for reviewing operations. Kavita Anand, technical staff member, product, at Hebbia, says the model knows how to generate accurate deliverables, consistent with the instructions received. She envisages hours saved thanks to a first version closer to the expected result, going so far as to mention a few minutes of proofreading instead of nights of work. These are anticipated gains, not a measure of time saved during testing.
Results on sources and presentations
To compare the models, Hebbia submitted identical financial cases to them with a common rating grid. Business experts established or verified the criteria based on the original parts. For the categories of sources to be cited, the company gives GPT-6 Astra a lead of 13.6 percentage points over the best alternative tested. It thus describes an expanded documentary coverage, associated with a strong relevance of the references retained.
On the presentations, Hebbia indicates that 97.9% of the applicable requirements were met, a gap of 14.4 percentage points with the best competing solution evaluated. The model comes, according to the company, in the lead both for finish and for compliance with requests. Kavita Anand says she has not previously observed this double first place: the other models rather stood out on one of these dimensions.
The evening letterFree · 7 p.m.
Are you following the race for AI models?
Each output (GPT, Claude, Gemini, Mistral, etc.) decrypted the same evening, in 5 minutes. Free.
Included upon registration: our selection of the best AI guides & comparisons.
Choose your pace
Free · No spam · Unsubscribe in 1 click
Kavita Anand wants slides that respect the company’s graphic identity, without the appearance of automated production. She describes the progress in precision and quality observed as surprising, while recalling the partnership maintained with OpenAI since the beginning of Hebbia. On the scale of a room containing thousands of rooms, she estimates that a percentage point gained can already bring many marginal benefits.
Matrix links financial assessment to supporting documents
Before advising an acquisition, a banker examines the accounts, contracts concluded with clients and management presentations. Matrix brings this research together by connecting the answers to the pieces that support them. In the process evaluated by Hebbia, GPT-6 Astra chooses documents or cells from Matrix to support an answer, then converts the analysis into native PowerPoint slides according to the requested organization and instructions.
Hebbia illustrates this use with a fictitious operation involving an industrial software publisher, with around 2,000 files to study. The banker would seek to assess the sustainability of revenues and compare the company to its peers. After importing the document room, it could examine customer concentration, contractual conditions and unsubscriptions in cells provided with references.
In this hypothetical example, Matrix could supplement the filing with comparable company filings and industry news to support the acquisition thesis. The research could then become a summary of diligence and an evaluation support using the company’s template. Kavita Anand points out that users make millions of dollars in decisions based on AI responses; it attributes to citations a role of traceability and confidence in their accuracy.
Kavita Anand considers financial intelligence officers
Kavita Anand ultimately imagines agents who would track information in the background and transmit signals to the professionals responsible for evaluating it. Receiving an alert a few hours before the competition would allow, in the scenario she evokes, to begin the examination earlier. She believes that this monitoring could uncover more opportunities than manual searching and provide value to users.
Hebbia wishes to extend the results obtained to support the evaluation of financial transactions with research whose supporting documents remain consultable.

