OpenAI on November 20 published a customer story1 in which Hebbia reports that GPT-6 Astra outscored other models in its tests of source citation and presentation building for deal research.

The figures are Hebbia's own, and the OpenAI page is the only place they originate. Banks and advisory firms that buy AI for diligence work have little else to compare, because the test set, the rival models and the margins of error are not disclosed.

What Hebbia tested

Hebbia's Matrix platform, according to the case study, brings together research from documents such as financial statements, customer contracts and management presentations, and links answers to the documents behind them. The case study says Hebbia has been a partner of OpenAI "over the course of Hebbia's lifetime." It does not describe GPT-6 Astra's architecture, size or pricing.

Hebbia tested a prerelease build in August-September 2026, using the same financial research cases and grading criteria for each model. Domain experts wrote or audited the criteria against source documents, the case study says. It measured two things: finding the sources a question requires, and building a presentation that follows the banker's brief.

On sourcing, GPT-6 Astra scored 13.6 percentage points above the best alternative tested on citing the required types of sources. Of the excerpts it cited, 99.5% were relevant to the question.

On presentations, it met 97.9% of applicable requirements, 14.4 percentage points ahead of the best alternative tested on that measure. "GPT-6 Astra is the first model we've seen where it's number one on both of those fronts," a Hebbia speaker identified in the case study only as Kavita says. The case study adds that other models traded off polish against adherence to the brief.

Limits of the claims

The case study does not name the alternative models, release the evaluation dataset or report whether the gaps are statistically significant, and it includes no independent test results. A TokenPost report2 repeats the same figures and quotes Kavita Anand, identified there as a member of technical staff at Hebbia, on a different point: "Citations help with traceability as well as giving users confidence that the AI found the right answer." The TokenPost piece shows no independent testing.

The case study's 2,000-file industrial-software acquisition is a hypothetical workflow, not a measured result. The measured results cover only the two stated metrics. The case study also reports no measured time savings, though Hebbia's speaker says users can "imagine the hours this will save."

OpenAI's GPT-6 Astra launch page could not be retrieved, and the case study does not say whether, when or on what terms the model is generally available, or whether Hebbia uses it in production.