Hebbia reports GPT-6 Astra outperforms alternatives in financial research tests
·2026.11.20 09:00
Key point
In tests of a prerelease build, GPT-6 Astra achieved 99.5% citation relevance and met 97.9% of presentation requirements, outperforming the best alternative by 14.4 points on presentation quality.
Details
- Hebbia tested GPT-6 Astra in August-September 2026 and found it leads the best alternative by 13.6 percentage points on source citation relevance and 14.4 percentage points on presentation requirements.
- The model achieved 99.5% citation relevance, with cited excerpts directly relevant to the research question, and met 97.9% of applicable presentation requirements.
- GPT-6 Astra ranked first on both instruction following and output quality, allowing Hebbia's Matrix to generate native PowerPoint slides that adhere to firm branding and structure.
- The workflow supports traceability by linking slide footnotes and citations back to specific source documents and highlighted passages for verification.
- Hebbia plans to use these results to help finance professionals reduce time spent on reformatting and source checking, potentially enabling background agents to surface market signals.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.