Hebbia put GPT-6 Astra to work inside Matrix to find required sources and draft client-ready decks, OpenAI says.
The pitch is fewer overnighters fixing citations and slides.
Kavita Anand, Member of Technical Staff for Product at Hebbia, said the model surprised the team on output quality and accuracy. In Hebbia tests, GPT-6 Astra met 97.9% of applicable presentation requirements, according to the company, and it led the best alternative tested by 14.4 percentage points on that measure. Anand said it was the first model the team saw rank first on both craft and following guidelines.
That matters because the job is traceable research, not chat. In the hypothetical deal cited by the company, a banker loads about 2,000 files on an industrial-software target into Matrix, then asks about revenue durability, customer concentration and peer comps. Matrix pulls documents or cells to back each answer, then builds native PowerPoint slides in the firm template with footnotes back to sources.
You have seen where this breaks. Slide text looks clean until you check the period or the cell.
Hebbia says citations are the fix. In its tests, GPT-6 Astra held a 13.6 percentage-point lead over the best alternative tested on citing required source types, according to the company, while 99.5% of cited excerpts were relevant to the question. Users can open a citation to see the highlighted passage, then correct details like the reporting period before a deck goes to committee.
Keep the test limits in view. Hebbia tested a prerelease build in August-September 2026 using the same financial research cases and grading criteria for each model, according to the company, with domain experts writing or auditing the criteria. Matrix itself runs an agent swarm architecture with multiple agents splitting complex work, per a third-party product breakdown, so the model gain sits inside Hebbia orchestration rather than as a solo model. Bankers still own the final check, and teams with no data room chaos can ignore this for now.





