🔍 Read the full analysis: GPT-6 Astra And Hebbia: A New Approach To Financial Research on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Hebbia says its tests of a prerelease GPT-6 Astra found better source coverage and presentation compliance than the strongest alternatives it evaluated. The company-run results used selected financial research cases and do not establish performance in live deals or other workflows.
Hebbia says a prerelease version of OpenAI’s GPT-6 Astra performed better than the strongest alternative it tested at selecting supporting sources and producing presentations that met a banker’s instructions. The company ran the evaluations in its Matrix financial research platform in August and September 2026, but the results are company-run tests, not evidence of performance across live transactions or other uses.
Hebbia tested two capabilities: choosing documents or passages to support an answer, and creating native PowerPoint slides that follow a specified structure and guidelines, a focus related to GPT-6 Astra’s hidden reasoning. The company says the models were evaluated using the same financial research cases and grading criteria. Domain experts wrote or audited the criteria against the underlying documents, according to Hebbia.
On source selection, GPT-6 Astra scored 13.6 percentage points above the best alternative tested for citing required types of sources. Hebbia also reported that 99.5% of cited excerpts were relevant to the question. For presentation tasks, the model met 97.9% of applicable requirements, 14.4 percentage points above the best alternative on that measure. The company did not provide the full set of model competitors or enough information to independently assess the scoring.
Hebbia illustrated the workflow with a hypothetical banker reviewing about 2,000 files for a potential acquisition, an application that echoes GPT-6 Astra’s end-to-end systems work. Matrix could help examine customer concentration, contract terms and churn, add outside material such as peer filings and industry news, and draft a diligence summary and valuation presentation. Hebbia says citations and slide footnotes can link back to source documents so a user can inspect the evidence. The example is not described as a completed client deal.
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.
Why Source Links and Decks Matter
The reported tasks reflect two distinct demands in financial research: assembling evidence from a large document set and communicating findings in a format that meets a firm’s requirements. Source coverage can affect whether an analyst’s answer rests on the relevant documents; presentation compliance can affect how much editing is needed before a draft is usable.
Hebbia’s results may be useful as an account of how one platform is evaluating AI for document-heavy finance work. They do not show that a generated answer is correct simply because it cites a relevant passage, or that a presentation is ready to use without review. A reviewer still needs to check whether a figure uses the right reporting period and whether the cited evidence supports the conclusion. The company reported no measured time savings, review-error rate, or effect on client decisions.
The distinction between those two scores also matters: a polished deck can omit key evidence, while a well-supported analysis can still require substantial work to turn into a client-ready presentation. Hebbia says GPT-6 Astra ranked first among the models it tested on both measures, but independent evaluations would be needed to judge how well that result generalizes.
Products Worth Considering

Amazon Kindle Scribe Colorsoft 64GB with Premium Pen

Samsung Galaxy Tab A11+ Kids Edition, 128GB, Royal Blue

Samsung 32 Odyssey G5 QHD Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How Matrix Handles Deal Research
Matrix is Hebbia’s platform for bringing research together from documents and linking answers to source material. In the company’s acquisition scenario, a user could work across data-room files and external information, then create a diligence summary and slides using a firm’s template. The cited passages and slide footnotes are intended to let professionals inspect the material behind generated claims.
The evaluation, as described by Hebbia, focused on source selection and slide generation rather than the full range of financial work. The company says it used consistent cases and criteria across models, with domain experts involved in developing or auditing the grading. That setup gives the reported comparisons a defined scope, but the available description does not establish how the results would change with different documents, instructions, or evaluation methods.
Hebbia also described a possible future role for agents that monitor information and surface signals for finance professionals. The source material provides no release date or test results for that proposed capability, so it should be treated as a forward-looking idea rather than a reported product outcome.
“The model scored above the best alternative Hebbia tested on source coverage and presentation requirements.”
— Hebbia
What the Tests Do Not Establish
The reported figures come from Hebbia’s own evaluation of a prerelease model. The material does not specify the number of test cases, the complete list of competing models, scoring variability, or how individual criteria were weighted. It is also unclear how the results would hold up with other financial research tasks or in live deal settings where files may be incomplete, inconsistent, or updated.
Hebbia did not report measured time savings, the frequency of errors found during human review, or outcomes for clients using the workflow. The example involving an acquisition is hypothetical, and the company’s suggestions that the tools could reduce overnight work or speed review are expectations rather than reported deployment results. No independent benchmark or external replication is described.
The proposed agent workflow also remains unspecified: Hebbia has not provided a timeline, detailed capabilities, or evaluation results for systems that would monitor information and flag developments. Readers should distinguish that future-facing concept from the tested source-selection and presentation tasks.
Evidence Needed Beyond Company Tests
The next useful evidence would include broader evaluations with more detail about cases, competing models, scoring methods, and variation in results. Independent testing could help show whether GPT-6 Astra’s reported advantage persists outside Hebbia’s selected tasks.
For practical adoption, results from actual workflows would clarify how professionals review generated answers and presentations, how often they correct errors, and whether the system measurably changes research time or decision processes. Hebbia has also described interest in agents that monitor information, but has not given a release date or reported performance for that work. Until more evidence is available, the findings remain a company-reported comparison of a prerelease model on defined financial research tests.
Key Questions
What did Hebbia test GPT-6 Astra on?
Hebbia evaluated source selection for supporting answers and the creation of PowerPoint presentations that follow specified structures and guidelines.
What results did Hebbia report?
The company said GPT-6 Astra scored 13.6 percentage points above the best alternative tested on required-source coverage and met 97.9% of applicable presentation requirements, 14.4 percentage points above the best alternative on that measure. It also reported that 99.5% of cited excerpts were relevant to the question.
Were the results independently verified?
The results described are from Hebbia’s company-run tests. The available account does not report an independent replication or provide enough detail to assess every part of the methodology.
Does this show GPT-6 Astra is ready to handle live deals?
No. The results do not establish performance in live transactions. Hebbia did not report client outcomes, measured time savings, or how often reviewers found errors. Generated analysis and citations would still need professional review.
When will Hebbia’s proposed monitoring agents be available?
Hebbia has described agents that could monitor information and surface signals, but the source material gives no release date, detailed capabilities, or test results for that proposed role.
Primary source: OpenAI · via ThorstenMeyerAI.com
