GPT-6 Astra improves financial research accuracy and presentation quality
Hebbia's testing shows the new model leads competitors in finding relevant sources and creating professional-grade slide decks for finance.
The short version
- GPT-6 Astra scored a 13.6 percentage-point lead over the best alternative in citing required source types.
- The model achieved a 99.5% relevance rate for the excerpts it cited during Hebbia's tests.
- It met 97.9% of presentation requirements, beating the next best model by 14.4 percentage points.
- The model can generate native PowerPoint slides that follow specific firm branding and structural guidelines.
Why it matters
For professionals in high-stakes fields like finance, AI is only useful if it is accurate and follows strict formatting rules. This update suggests a shift toward AI that can handle complex, multi-step workflows—from research to final presentation—rather than just generating text.
Testing the new model
Hebbia recently conducted a series of tests to evaluate the capabilities of OpenAI’s GPT-6 Astra. The company focused on two specific areas of high-value work for finance professionals: identifying the correct sources for a question and building a presentation that follows a specific brief. These tests were conducted using a prerelease build of the model between August and September 2026.
The research was designed to mimic real-world scenarios, such as a banker assessing a strategic acquisition of an industrial-software company. In this hypothetical case, the user would need to review roughly 2,000 files to understand revenue durability, peer comparisons, and customer concentration. Domain experts wrote and audited the grading criteria against actual source documents to ensure the results were meaningful.
The results showed that GPT-6 Astra provided a significant advantage over other models. As Kavita from Hebbia noted, the leaps the model made in quality and accuracy were surprising. The company found that the model was particularly strong at combining broad source coverage with highly relevant citations, which is critical for making multi-million dollar decisions.
Accuracy and traceability
One of the primary challenges in using AI for financial research is the need for traceability. Because bankers base major decisions on the information provided by the AI, they must be able to verify every claim. Hebbia’s Matrix platform helps by linking every answer directly to the underlying documents, such as customer contracts or management presentations.
GPT-6 Astra demonstrated high precision in this area, scoring a 13.6 percentage-point lead over the best alternative for citing the required source types. Furthermore, 99.5% of the excerpts the model cited were found to be relevant to the specific questions asked. This level of accuracy helps users feel confident that the AI is finding the correct data points within a massive data room.
The platform allows users to open a citation to see the specific document and the highlighted passage. This enables a researcher to quickly check a figure against a source table or correct a reporting period before the information reaches an investment committee. By providing these links, the model helps ensure that the final output is grounded in verifiable evidence rather than AI-generated hallucinations.
Meeting presentation standards
Beyond just finding facts, the model was tested on its ability to create professional-grade artifacts. Finance professionals often require outputs to be in a specific firm's branding and to follow a human-made structure. Hebbia found that GPT-6 Astra met 97.9% of applicable presentation requirements, a 14.4 percentage-point lead over the next best model tested.
In many previous tests, models often traded off polish for instruction following. A model might be good at creating a beautiful slide but fail to follow the specific guidelines provided by the user. GPT-6 Astra was the first model in Hebbia's tests to rank first on both dimensions, successfully adhering to the user's brief while maintaining high-quality craft.
The model can turn research into native PowerPoint slides that follow a firm's specific template. This means it can take analysis on peer filings and industry news and format it into a diligence summary. This capability is intended to help professionals avoid the tedious task of reformatting decks, allowing them to focus more on the actual assessment of a deal.
Impact on professional workflows
The goal of these improvements is to significantly reduce the time spent on manual
Reducing manual labor for bankers
The improvements in model accuracy are significant because even a single percentage-point gain can provide substantial marginal benefits when reviewing thousands of documents. By creating precise outputs that adhere strictly to user guidelines, the model can produce first drafts that are much closer to a final product. This allows professionals to spend less time on manual formatting and more time on high-level deal assessment.
This shift in workflow could fundamentally change how teams handle demanding projects. Kavita from Hebbia suggests that the long "overnighters" typically required to build these decks could be reduced to just a few minutes of reviewing the AI-generated output. By automating the heavy lifting of research and drafting, the technology aims to let humans focus on the strategic thinking that requires human judgment.
Future of automated signals
Looking ahead, Hebbia envisions a role for AI agents that monitor information in the background to surface important signals for finance professionals. These agents could potentially identify key data points a few hours before competitors do, providing a significant competitive advantage. This proactive approach allows firms to stay ahead of market trends without having to manually hunt for every piece of information.
By surfacing these signals automatically, the technology can help professionals see and process more opportunities than a human could find alone. Instead of manually searching through vast amounts of data, users can get started thinking about new deals sooner. This transition from manual hunting to automated signal processing is intended to unlock more value for users by highlighting opportunities that might otherwise be missed.
Sources and links
- ReportingOpenAI: the article we sourced
First reported by OpenAI. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.