OpenAI says Hebbia saw better financial research results with GPT-6 Astra
OpenAI’s Hebbia customer story says a prerelease GPT-6 Astra build improved source coverage, citation relevance and presentation quality in financial-research tests run in August and September 2026.

- OpenAI published a customer story on 20 November 2026 about Hebbia’s tests of GPT-6 Astra.
- Hebbia says it used the same financial-research cases and grading criteria across models, with domain experts writing or auditing the criteria against source documents.
- The story says GPT-6 Astra led the best alternative by 13.6 percentage points on required source types and 14.4 percentage points on presentation requirements.
- OpenAI says 99.5% of cited excerpts were relevant to the question and 97.9% of applicable presentation requirements were met in Hebbia’s tests.
OpenAI has published a customer story saying Hebbia saw stronger financial-research results with a prerelease build of GPT-6 Astra. The story is dated 20 November 2026 and describes tests run by Hebbia in August and September 2026. OpenAI customer story
What Hebbia tested
According to OpenAI’s write-up, Hebbia ran the same financial-research cases and grading criteria across models. Domain experts wrote or audited those criteria against the source documents, which matters because it anchors the results in the documents bankers and analysts actually use. OpenAI customer story
The numbers OpenAI published
| Measure | GPT-6 Astra result | What OpenAI says it led by |
|---|---|---|
| Required source types cited | 13.6 percentage-point lead over the best alternative | 13.6 points |
| Citation relevance | 99.5% of cited excerpts were relevant to the question | Not stated |
| Presentation requirements met | 97.9% met | 14.4 percentage-point lead over the best alternative |
OpenAI says GPT-6 Astra combined broader source coverage with highly relevant citations. It also says the model met 97.9% of applicable presentation requirements, which suggests the output formatting and structure were part of the evaluation, not just answer accuracy. OpenAI customer story
Why this matters for developers and research teams
For teams building document-heavy workflows, the story is less about a benchmark headline than about evaluation design. Hebbia’s setup used the same cases and criteria for each model, with human experts checking the rubric against the underlying documents. That makes the published numbers more useful than a generic model claim, even though they still come from a vendor-published customer story. OpenAI customer story
The practical takeaway is straightforward: if your application depends on source grounding, citation quality and presentable output, OpenAI is positioning GPT-6 Astra as an option to assess against those constraints rather than only raw answer quality. OpenAI customer story
What is still unknown
The story does not provide the full benchmark set, the exact competing models, or the scoring rubric in detail. It also does not show how the prerelease build performs in production beyond Hebbia’s own tests. Readers should treat the published figures as official customer-story results, not independent verification. OpenAI customer story
- Required source types cited lead13.6 percentage points / percent
- Presentation requirements lead14.4 percentage points / percent
- Citation relevance99.5 percentage points / percent
- Presentation requirements met97.9 percentage points / percent

