Skip to content
trendoviz

Hebbia tests GPT-6 Astra on financial research

Hebbia evaluated a prerelease build of OpenAI’s GPT-6 Astra on source selection and presentation generation within its Matrix financial-research platform. The company reported leads over the best alternatives it tested, but the underlying cases, comparison models and scoring methods were not published.

52 out of 100Emerging7D: Insufficient history ProvisionalPublished

Overview

Hebbia is a financial-research platform whose Matrix product links answers to supporting documents. In the evaluated workflow, GPT-6 Astra was used to identify relevant source material and produce native PowerPoint slides that follow a user’s structure and guidelines [1].

The evaluation took place in August–September 2026 and used the same financial-research cases and grading criteria for every model. Domain experts wrote or audited the criteria against source documents [1]. Hebbia reported that GPT-6 Astra covered broader sources, produced more relevant citations and followed presentation requirements more closely than the alternatives it tested [1].

These are results reported by Hebbia in an OpenAI News article. The article does not identify the tested alternatives or provide complete per-model results, so the comparisons cannot be independently reconstructed from the published figures [1].

Why it matters

The reported findings concern workflow output rather than general financial accuracy. A model that can retrieve supporting material, cite it and format the result as a presentation could reduce manual work between research, verification and deck preparation, leaving finance professionals more time to assess the analysis [1].

Traceability is particularly relevant in this setting. Kavita said users base millions of dollars of decisions on AI-generated answers, and that citations support both traceability and confidence [1]. The practical implication is a human-checked research process in which reviewers can inspect linked evidence rather than accepting a polished answer on its appearance alone [1].

Because the published evaluation covers only part of the platform’s activity, it does not establish that GPT-6 Astra improves an entire end-to-end deal workflow [1].

Opportunities

  • Traceable due-diligence drafting: Teams could require reviewers to verify a figure or claim against its linked source before approving a generated summary or presentation [1].
  • Template-aware financial presentations: Hebbia describes native PowerPoint output that follows a user’s structure and guidelines, which could support first drafts of diligence materials and valuation decks that still require professional review [1].
  • Evidence-selection services: Developers could build research assistants that retrieve relevant filings, contracts and presentations while preserving links to the material behind each answer [1].
  • Monitoring agents for finance: Kavita described a possible future in which agents watch for deal or market signals and surface them for assessment. She presented this as an imagined capability rather than a released feature [1].
  • Internal model evaluation: Organizations could define source-coverage, citation-relevance and presentation-adherence criteria and compare models on recurring internal research cases [1].

Risks

  • Incomplete disclosure: The article names no alternative models and provides no raw scores or full per-model results, leaving the size and robustness of the reported leads difficult to assess [1].
  • Prerelease versus released behavior: The evaluation used a prerelease build, and the published findings do not establish that the generally released model produces the same results [1].
  • Unclear scoring method: The article does not explain whether the citation-relevance measure came from human review, automated scoring or another method [1].
  • Workflow reliance: A system that searches, cites and formats at once can amplify retrieval or citation errors if reviewers trust a polished result without checking the underlying evidence [1].
  • Hypothetical scale: Hebbia’s acquisition-review example involved roughly 2,000 files but was explicitly hypothetical, so it provides no measured evidence of performance at that scale [1].

Context

Financial research typically separates into two connected activities: finding the documents relevant to a question and turning that research into a deliverable that colleagues or clients can review. Hebbia tested GPT-6 Astra on both activities inside Matrix, covering source identification and presentation generation [1].

The workflow is designed around source-level traceability. Hebbia says users can open a citation to view the underlying document and highlighted passage, while slide footnotes also lead back to source material [1]. This allows a reviewer, for example, to check a customer-concentration figure against its source table and correct a reporting period before a deck proceeds [1].

Known, claimed, inferred

Reported findings: GPT-6 Astra scored 13.6 percentage points above the best alternative tested on citing the required types of sources. Hebbia also reported that 99.5% of the excerpts it cited were relevant to the question [1]. On presentation work, the model met 97.9% of applicable requirements, a lead of 14.4 percentage points over the best alternative tested on that measure [1].

Vendor claims: Hebbia says GPT-6 Astra selects documents or Matrix cells to support an answer and converts the analysis into native PowerPoint slides that follow the user’s structure and guidelines [1]. It also says citations open to the underlying document and highlighted passage, and that slide footnotes link to source material [1].

Testimonial: Kavita said GPT-6 Astra was the first model Hebbia had seen rank first on both presentation craft and adherence to guidelines [1].

Hypothesis: The results suggest a combined advantage in evidence retrieval and deliverable formatting. That could reduce handoffs between research, checking and presentation work, but the findings do not establish accuracy in financial judgment or performance across the full Matrix workflow [1].

Limitations & uncertainties

  • The evidence consists of two items from the same OpenAI News article rather than an independent third-party test report [1].
  • The number of financial-research cases, the evaluation sample size and any statistical-significance analysis are not disclosed [1].
  • The financial-research cases and domain-expert grading criteria are not published, and the distinction between source coverage and required-source-type coverage is not explained [1].
  • The applicable presentation requirements are unspecified, as is whether every evaluated model produced native PowerPoint slides [1].
  • It is unclear whether the evaluation exercised the entire Matrix workflow or only source selection and presentation generation [1].
  • The proposed role for monitoring agents was presented as a future possibility rather than a demonstrated capability [1].

Evidence & sources

The key statements of this analysis, labeled by type, with the sources they rely on. Trendoviz does not run product tests.

  • Primary-source or vendor statementHebbia tested a prerelease build of GPT-6 Astra on source identification and presentation generation within its Matrix financial-research platform.[1]
  • Primary-source or vendor statementThe evaluation used the same financial-research cases and grading criteria for each model, with domain experts writing or auditing the criteria against source documents.[1]
  • Primary-source or vendor statementHebbia reported that GPT-6 Astra scored 13.6 percentage points above the best alternative tested on citing required source types, with 99.5% of its cited excerpts judged relevant to the question.[1]
  • Primary-source or vendor statementHebbia reported that GPT-6 Astra met 97.9% of applicable presentation requirements, placing it 14.4 percentage points ahead of the best alternative tested on that measure.[1]
  • User testimonialKavita said GPT-6 Astra was the first model Hebbia had seen rank first on both presentation craft and adherence to guidelines.[1]
  • Primary-source or vendor statementHebbia says GPT-6 Astra selects documents or Matrix cells to support answers and turns analysis into native PowerPoint slides that follow a user's structure and guidelines.[1]
  • Primary-source or vendor statementThe acquisition-review example involving roughly 2,000 files was explicitly hypothetical and was not a measured deployment result.[1]
  • User testimonialKavita described a possible future role for agents that monitor information and surface signals for finance professionals to assess.[1]

Sources

  1. 1

Updates & corrections

  1. First published · published automatically after passing every check

Found an error? Report it.