TL;DR
Hebbia says its tests of a prerelease GPT-6 Astra found gains in finding relevant sources and producing presentations that follow a banker’s brief. The results come from company-run tests using financial research cases; they do not establish how the model performs across other workflows or in live deals.
Hebbia says a prerelease build of OpenAI’s GPT-6 Astra improved two parts of its financial research workflow: finding the sources needed to support an answer and creating presentations that follow a user’s instructions. In tests conducted in August and September 2026, the model scored above the best alternative Hebbia tested on source coverage and presentation requirements, according to the company.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Hebbia tested the model in Matrix, its platform for bringing research together from documents and linking answers back to their sources. The company used the same financial research cases and grading criteria for each model. Domain experts wrote or audited those criteria against the underlying documents, Hebbia said.
On source selection, GPT-6 Astra scored 13.6 percentage points above the best alternative tested at citing required types of sources. Hebbia also reported that 99.5% of cited excerpts were relevant to the question. On presentation tasks, the model met 97.9% of applicable requirements, 14.4 percentage points above the best alternative on that measure.
The company’s example describes a banker reviewing roughly 2,000 files for a potential acquisition. Matrix could help examine customer concentration, contract terms and churn, draw on peer filings and industry news, and create a diligence summary and valuation presentation. Hebbia says citations can link answers and slide footnotes back to documents, which lets users inspect supporting passages and figures.
Hebbia reports GPT-6 Astra gains in financial research
Company-run tests found improvements in source selection and presentation requirements. The results cover selected research cases, not live deals or other workflows.
What Hebbia tested
Tests in August and September 2026 used financial research cases and the same grading criteria for each model.
Find supporting evidence
GPT-6 Astra scored higher on citing required types of sources. Matrix links answers to documents for review.
Follow the banker’s brief
The model produced native PowerPoint slides measured against specified structure and guidelines.
From files to a deal summary
Hebbia’s hypothetical acquisition example shows how document research could feed a presentation.
Examine roughly 2,000 files for details such as customer concentration, contract terms and churn.
Bring in peer filings and industry news alongside the data room documents.
Create a diligence summary and valuation slides, then inspect citations and footnotes against their sources.
What the results establish
The reported scores describe performance under Hebbia’s evaluation.
- Domain experts wrote or audited the criteria.
- GPT-6 Astra ranked first among tested models on both measures.
- How results transfer to messy, changing live deal documents.
- Actual time saved, review error rates and independent replication.
Citations and Decks in Deal Work
The reported gains target tasks where mistakes or extra review can slow down financial work. A banker evaluating a deal may need to connect figures and claims across many files, then explain the analysis in a presentation that meets firm requirements. Relevant citations can help reviewers check where an answer came from, while a closer-to-brief first draft could reduce some formatting and editing work.
Hebbia’s findings matter as an account of how one financial research platform is testing AI for document-heavy work. The figures describe performance against the company’s selected cases and criteria; they do not show that bankers can rely on generated analysis without review. Users still need to verify details such as the reporting period behind a customer-concentration figure and whether a cited passage supports the conclusion drawn from it.
The distinction between source coverage and presentation quality is also practical. A polished deck is not useful if it misses required evidence, and a well-supported answer may still take substantial work to turn into a client-ready presentation. Hebbia says GPT-6 Astra ranked first among the models it tested on both dimensions, but independent results would be needed to judge whether that balance holds more broadly.
Products Worth Considering

Samsung 27-Inch Odyssey G5 QHD Gaming Monitor (G50F)

Samsung Galaxy Tab A11+ Tablet (6GB RAM, 128GB Storage, Gray)

Sonos Five Wireless HiFi Speaker, Black
As an affiliate, we earn on qualifying purchases.
How Matrix Connects Research
Matrix is designed to bring together information from files such as financial statements, contracts and management presentations, with answers tied to supporting documents. In the acquisition example described by Hebbia, a user could review documents in a data room and add outside material such as peer filings and industry news. The platform could then help turn that work into a diligence summary and slides using the firm’s template.
The workflow places traceability alongside drafting. A user can open a citation to inspect the document and highlighted passage, and slide footnotes can link back to source material. Hebbia presents this as a way for professionals to check the evidence behind an answer before using it in a decision process.
The tests focused on two model capabilities: selecting documents or Matrix cells to support an answer, and producing native PowerPoint slides that follow a specified structure and guidelines. The example of a banker assessing an industrial-software company is hypothetical; Hebbia did not describe it as a completed client deal.
Limits of the Reported Tests
The reported results come from Hebbia’s own tests of a prerelease model build. The company has not provided enough detail here to assess the number of cases, the full set of competing models, scoring variability or how the criteria were weighted. The test figures therefore show how the models performed under Hebbia’s evaluation, not a general measure of accuracy in financial research.
It is also unclear how the reported results translate to live deal work, where documents may be incomplete, inconsistent or updated over time. Hebbia did not quantify actual time saved, report the rate of errors found in review, or describe whether clients have measured effects on decisions. Its statements about reduced overnight work and faster review are expectations, not reported deployment outcomes.
The company also described a future role for agents that monitor information and surface signals to finance professionals. No deployment timeline, capabilities or evaluation results for that proposed role were provided.
From Test Results to Use
Hebbia says it wants to build on the reported results to help finance professionals assess deals using research they can trace back to its sources. The next evidence that would clarify the practical value would include results from broader evaluations and details on how users review generated answers and presentations in actual work.
The company also sees potential for agents to monitor information in the background and flag developments for finance professionals to assess. For now, that is a forward-looking idea described by Anand, rather than a feature with reported performance or a stated release date.
Where I land
I see Hebbia’s results as a useful, narrow signal: source selection and presentation instructions are concrete tasks that matter in financial research, and the reported margins give a reason to examine the model further. The citation figures are most persuasive when they help a reviewer quickly verify a claim against the right document and reporting period.
The strongest counterargument is that company-run tests may not reflect messy live deal work, and a highly relevant excerpt can still be used incorrectly. Hebbia has not reported independent replication, real-world error rates or measured time savings. I would raise my assessment if external evaluations across varied data rooms showed similar gains and deployments documented faster review without weakening source checks.
Source: OpenAI
Key Questions
What did Hebbia test GPT-6 Astra on?
Hebbia tested how well the model selected sources for financial research answers and produced presentations that followed a user’s brief.
What results did Hebbia report?
The company said GPT-6 Astra scored 13.6 percentage points above the best alternative tested on citing required source types, and met 97.9% of applicable presentation requirements, 14.4 percentage points above the best alternative on that measure.
Were the results independently verified?
The results described here are from Hebbia’s tests. The company said domain experts wrote or audited the grading criteria, but no independent evaluation is reported.
Can users check the evidence behind generated answers?
Hebbia says Matrix citations open the underlying document and highlighted passage, while slide footnotes can link back to source material. Users still need to review whether that evidence supports the answer.
Source: OpenAI
