Research

AI-generated text

Hebbia reports financial research gains in GPT-6 Astra tests

Hebbia says OpenAI’s GPT-6 Astra performed well in tests of source selection and presentation creation for financial research.

Hebbia reports financial research gains in GPT-6 Astra tests

Hebbia says tests of OpenAI’s GPT-6 Astra showed improvements in two important parts of financial research: selecting sources needed to support answers and creating presentations that follow a banker’s brief.

Hebbia’s Matrix product brings financial research materials together and links answers to the documents behind them. That can help when a banker must assess financial statements, customer contracts and management presentations before recommending an acquisition.

In a hypothetical example from Hebbia, a banker is assessing the strategic acquisition of an industrial-software company with roughly 2,000 files to review. With the data room loaded into Matrix, the banker could examine customer concentration, contract terms and churn in cells backed by citations. The system could also bring in peer-company filings and industry news to inform the deal assessment, then turn the research into a diligence summary and valuation deck using the firm’s template.

In that workflow, GPT-6 Astra selects documents or Matrix cells to support an answer, then turns the analysis into PowerPoint slides that follow the user’s structure and guidelines. Kavita, speaking for Hebbia, said traceability matters because users may base decisions worth millions of dollars on the AI’s responses.

Testing sources and presentations

Hebbia tested a prerelease version of the model in August and September 2026. Each model in the comparison received the same financial research cases and grading criteria, which domain experts wrote or audited against the source documents.

The company said GPT-6 Astra scored 13.6 percentage points above the best alternative tested on whether it cited the required types of sources. Of the excerpts it cited, 99.5% were relevant to the question. Kavita said that when thousands of documents are involved, even a one-percentage-point gain can add substantial value.

For presentations, the model met 97.9% of applicable requirements, a lead of 14.4 percentage points over the best alternative tested on that measure. Hebbia said GPT-6 Astra ranked first in both following the brief and presentation quality, while other models showed trade-offs between polish and adherence to instructions. Kavita said bankers want slides that follow their firm’s branding and look like human-made work.

Research that can be checked

In Hebbia, users can open citations to see the original document and highlighted passage; slide footnotes also lead back to source material. A reviewer could, for example, compare a customer-concentration figure with the original table and correct the reporting period if necessary before a presentation reaches an investment committee.

Hebbia aims to help bankers spend less time reformatting presentations and checking sources, and more time assessing deals. Kavita also outlined a potential future role for agents that monitor information in the background and surface signals for finance professionals. Earlier signals, she said, could allow users to consider more opportunities while keeping the research behind their decisions traceable to its sources.