🔍 Read the full analysis: Hebbia’s Financial Research Gains With GPT-6 Astra 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 a prerelease build of OpenAI’s GPT-6 Astra outscored the best alternative it tested on source coverage and presentation requirements in financial research cases. The August and September 2026 tests were run by Hebbia and do not establish performance in live deals or other workflows.
As detailed in the original analysis, Hebbia says a prerelease version of OpenAI’s GPT-6 Astra outperformed the best alternative it tested on two financial research tasks: finding sources that support an answer and producing presentations that follow a banker’s instructions. The company-run tests, conducted in August and September 2026 in its Matrix platform, offer evidence about performance on Hebbia’s selected cases, not a general measure of accuracy or results in live deals.
On source selection, GPT-6 Astra scored 13.6 percentage points above the best alternative tested at citing required types of sources, a finding related to what Astra caught in Claude’s model edit. Hebbia also reported that 99.5% of the excerpts it cited were relevant to the question. The company said the models were tested on the same financial research cases and graded against the same criteria, which domain experts wrote or audited against the underlying documents.
For presentation tasks, Hebbia said Astra met 97.9% of applicable requirements, 14.4 percentage points above the best alternative on that measure. The tasks included selecting supporting documents or Matrix cells and producing native PowerPoint slides that followed specified structures and guidelines, similar to the end-to-end systems Perplexity trusts Astra to handle. Hebbia said Astra ranked first among the models it tested on both dimensions.
The company’s illustrative scenario involves 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, then prepare a diligence summary and valuation presentation. Hebbia says citations can link answers and slide footnotes to source documents. It described the scenario as hypothetical, not 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.
Evidence and Slides in Deal Review
The reported results address two connected demands in financial research: finding supporting evidence across large document sets and turning analysis into a presentation that follows a brief. In deal work, those steps can involve reviewing contracts, financial statements and management materials, then explaining findings in a format colleagues or clients can inspect. A useful improvement in either task could reduce some searching or editing, while a combination of the two could make a first draft easier to review.
But the measures do not establish that generated analysis can be used without professional checks. A relevant excerpt may still fail to support the conclusion drawn from it, and a figure can be misleading if its reporting period or definition is wrong. Reviewers would still need to verify the source, interpretation and context behind the claims before relying on them in a deal decision.
The distinction between source coverage and presentation quality matters, too. A polished deck can omit required evidence; a well-supported answer can still take substantial work to turn into a client-ready presentation. Hebbia’s results suggest the tested model performed well on both of its chosen measures, but the company has not reported whether that translates to less review time, fewer errors or better decisions in actual use.
Products Worth Considering

Amazon Kindle Scribe (16GB) with Premium Pen, Tungsten

Apple AirPods Pro (1st Generation) with MagSafe Charging Case

Jackery HomePower 3000 Portable Power Station, 3072Wh, 3600W
As an affiliate, we earn on qualifying purchases.
How Matrix Links Research Sources
Matrix is Hebbia’s platform for bringing research together from documents and linking answers back to supporting material. The workflow described by the company combines files such as financial statements, contracts and management presentations with outside material, including peer-company filings and industry news. Users can inspect a citation to open the source document and highlighted passage; slide footnotes can also link back to sources.
That traceability is part of the product’s intended role in professional research: users can check where an answer came from rather than treat generated text as a standalone conclusion. Hebbia says it used the same financial research cases and grading criteria for each model in its tests. However, the source material does not specify the number of cases, the full competitor list, scoring variability or the criteria’s weighting, details needed to judge how broadly the results apply.
The reported figures concern two tasks, not every part of financial analysis. They do not measure whether the system identifies every relevant document, interprets conflicting information correctly, or remains reliable when a data room changes. The acquisition example is an illustration of a possible workflow, rather than evidence of a customer outcome.
“GPT-6 Astra ranked first among the models it tested on both source coverage and presentation requirements.”
— Hebbia
Limits of Hebbia’s Test Results
The results come from Hebbia-run tests of a prerelease model. The available account does not provide the case count, complete list of competing models, score variability or details of how criteria were weighted. Without those details or independent evaluations, it is not possible to determine whether the reported margins would hold across different data, firms or research tasks.
Hebbia did not report measured time savings, error rates found during review or effects on client decisions. It is also unclear how the results would carry over to live deals, where documents can be incomplete, inconsistent or updated over time. The company’s expectations that the workflow could reduce overnight work or speed review are not reported deployment outcomes.
Hebbia has also described a possible future role for agents that monitor information and surface signals to finance professionals. The company has not provided a release date, detailed capabilities or evaluation results for that proposed use. The model’s broader availability and performance beyond the tested Matrix tasks are not established by the reported results.
Evidence Needed From Live Work
The next useful evidence would include broader evaluations with fuller methods, including the number and type of cases, competitor models, scoring rules and variability across runs. Results from independent testers could help show whether the measured gains extend beyond Hebbia’s chosen examples.
For practical use, readers would also need results from real workflows: how professionals check generated answers and slides, what errors they find, and whether the system changes research or review time. Hebbia has not reported those outcomes in the material available here. Its proposed information-monitoring agents remain a forward-looking idea, with no stated deployment timeline or performance data.
Key Questions
What did Hebbia test GPT-6 Astra on?
Hebbia tested a prerelease build in Matrix on financial research tasks involving source selection and creating presentations that follow specified requirements.
What results did Hebbia report?
The company said Astra scored 13.6 percentage points above the best alternative on citing required source types, with 99.5% of cited excerpts relevant to the question. It said Astra met 97.9% of applicable presentation requirements, 14.4 percentage points above the best alternative on that measure.
Were these independent tests or live deal results?
No. They were company-run tests using Hebbia’s selected research cases and criteria. The acquisition scenario was hypothetical, and Hebbia did not report outcomes from a completed client deal.
Do the results show that generated financial research needs no review?
No. The reported scores do not establish that answers are error-free or suitable for use without checking. Professionals would still need to verify citations, figures, time periods and the reasoning behind conclusions.
What remains unknown about GPT-6 Astra’s use in finance?
Hebbia has not supplied enough detail here to assess the number of test cases, the full competitor set or scoring variability. It also has not reported measured time savings, review error rates or decision outcomes from live work.
Primary source: OpenAI · via ThorstenMeyerAI.com
