Hebbia tests GPT-6 Astra: 99.5% relevant citations

Hebbia announces that 99.5% of the excerpts cited by GPT-6 Astra were relevant during its trials conducted in August-September 2026 on a preliminary version of OpenAI's model. The company evaluated the search for financial sources and the production of presentations that comply with guidelines. The time savings mentioned by Hebbia remain prospective.
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 review could, for example, focus on a customer concentration data compared to the original table, followed by the correction of the reporting period before submission to the investment committee.
The company aims to lighten the tasks of formatting and document verification to allow more time for reviewing operations. Kavita Anand, a technical staff member at Hebbia, states that the model can generate precise deliverables that comply with the instructions received. She envisions hours saved thanks to an initial version closer to the expected result, even suggesting a few minutes of proofreading instead of nights of work. These are anticipated gains, not a measure of time saved during the trials.
Results on Sources and Presentations
To compare the models, Hebbia submitted identical financial cases to them with a common scoring grid. Industry experts established or verified the criteria based on the original documents. 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 describes an expanded documentary coverage, associated with a high relevance of the selected references.
Regarding presentations, Hebbia indicates that 97.9% of applicable requirements were met, representing a gap of 14.4 percentage points with the best competing solution evaluated. According to the company, the model ranks first in both finishing and compliance with requests. Kavita Anand notes that she has not previously observed this dual first place: other models tended to excel in one of these dimensions.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Kavita Anand desires slides that respect the company's graphic identity, without appearing automated. She finds the improvements in accuracy and quality surprising, while recalling the partnership maintained with OpenAI since the inception of Hebbia. On the scale of a room containing thousands of documents, she believes that a one percentage point gain can already bring numerous marginal benefits.
Matrix Links Financial Evaluation to Supporting Documents
Before advising on an acquisition, a banker examines accounts, contracts with clients, and management presentations. Matrix consolidates this research by linking responses to the documents that support them. In the process evaluated by Hebbia, GPT-6 Astra selects documents or cells from Matrix to justify a response, then converts the analysis into native PowerPoint slides according to the requested organization and guidelines.
Hebbia illustrates this use with a fictional operation concerning an industrial software publisher, with about 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, he could examine customer concentration, contractual terms, and cancellations in cells with references.
In this hypothetical example, Matrix could complement the file with filings from comparable companies and industry news to support the acquisition thesis. The research could then become a synthesis of due diligence and an evaluation support document following the company's template. Kavita Anand emphasizes that users make decisions involving millions of dollars based on the AI's responses; she assigns citations a role in traceability and trust in their accuracy.
Kavita Anand Envisions Financial Monitoring Agents
Kavita Anand envisions eventually having agents that would monitor information in the background and send signals to professionals tasked with evaluating them. Receiving an alert a few hours before the competition would allow, in the scenario she describes, for an earlier start to the review. She believes that this monitoring could uncover more opportunities than manual research and provide value to users.
Hebbia aims to extend the results obtained to support the evaluation of financial operations with research whose supporting documents remain accessible.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.