Performance Metrics of GPT-6 Astra
Hebbia evaluated a prerelease build of GPT-6 Astra in August–September 2026 and reported measurable gains when the model was integrated into its Matrix workflow for financial research. The company focused on two primary tasks: identifying necessary documents for specific questions and producing presentation slides that adhere to a banker’s brief.
According to Hebbia’s tests, GPT-6 Astra outperformed alternatives on several key metrics. The model achieved a 13.6 percentage-point lead over the best alternative in citing the required types of sources. Furthermore, 99.5% of the excerpts cited by the model were relevant to the questions posed. In terms of presentation quality and adherence to user instructions, GPT-6 Astra met 97.9% of applicable presentation requirements, placing it 14.4 percentage points ahead of the best alternative tested. Domain experts wrote or audited the grading criteria against source documents, ensuring consistency across models.
Workflow and Traceability
Hebbia described a typical workflow where a banker loads a large data room—approximately 2,000 files in the example—into Matrix. GPT-6 Astra selects documents or individual Matrix cells to support answers, highlighting customer concentration, contract terms, and churn in cited cells. It then generates native PowerPoint slides that align with the user’s structure and branding guidelines.
“Traceability is really important to our users because they’re basing millions of dollars of decisions on what the AI actually comes back with,” said Kavita Anand, Member of Technical Staff, Product at Hebbia. The platform includes features that allow users to open citations to view the underlying documents and highlighted passages, with slide footnotes linking back to the source material for review and correction.
Improvements in Quality and Accuracy
Kavita Anand noted the surprising improvements in quality outputs and accuracy with GPT-6 Astra compared to previous models. She emphasized that GPT-6 Astra excels in both following user instructions and maintaining presentation quality, stating, “In the past, we’ve seen models that really shine on one or the other of these dimensions. GPT-6 Astra is the first model we’ve seen where it’s number one on both of those fronts.”
Future Developments and Agent Use
Looking ahead, Kavita Anand outlined a potential role for background agents that monitor information and surface signals for finance professionals. Hebbia aims to build on these results to assist finance teams in assessing deals with research that can be traced back to its sources, enhancing the decision-making process for users.
Original source: OpenAI News