Skip to content
Theme
OnDevice Tools

AIAI Models & Research

GPT-6 Astra financial research shows gains in Hebbia tests

Hebbia tested a prerelease build of OpenAI's GPT-6 Astra on financial research tasks and reported gains in source coverage, citation relevance and presentation quality. The results come from Hebbia's own evaluations, as described in an OpenAI News customer story.

Illustration of a banker reviewing AI-generated financial research slides with cited source documents

Key takeaways

  • Hebbia tested a prerelease GPT-6 Astra build in August and September 2026.
  • The model led the strongest alternative by 13.6 percentage points on citing required source types.
  • 99.5% of cited excerpts were relevant, and 97.9% of applicable slide requirements were met.
  • Hebbia says the results could reduce time spent checking sources and reformatting decks.
In this article
  1. What Hebbia measured
  2. Citations and source coverage
  3. Slides that follow the brief
  4. What it means for deal work

What Hebbia measured

Hebbia's GPT-6 Astra financial research tests looked at two parts of a banker's workflow: finding the documents needed to answer a question and turning that work into a presentation that follows a brief. The company used the same cases and grading rules for each model it tested, according to OpenAI News. Domain experts wrote or checked those rules against the source documents.

The tests ran on a prerelease build between August and September 2026. Hebbia makes Matrix, a research tool that links answers back to the documents behind them. In the customer story, OpenAI News says the model selects documents or Matrix cells to support an answer and then builds PowerPoint slides in the user's structure and house style.

Citations and source coverage

Illustration: AI, OpenAI logo, news, price and the labels Financial Research, Hebbia, OpenAI

On citing the types of sources a question required, GPT-6 Astra finished 13.6 percentage points ahead of the strongest competing model in Hebbia's tests. Almost all of the excerpts it cited, 99.5%, were relevant to the question. Those figures matter in finance because a single answer may rest on customer contracts, financial statements or management presentations.

Kavita, a Hebbia executive quoted in the customer story, said traceability is important when users base decisions worth millions of dollars on an AI answer. Citations let users open the underlying document and see the highlighted passage. In a data room with thousands of files, she said, even a small gain in accuracy can produce a meaningful benefit.

Slides that follow the brief

GPT-6 Astra also led on presentation quality. It met 97.9% of the applicable presentation requirements in Hebbia's tests, 14.4 percentage points ahead of the best alternative. Hebbia said other models tended to trade polish against adherence to instructions, while GPT-6 Astra ranked first on both measures.

Kavita said a deck should look as if it came from the banker's firm, not as if a machine made it. In Hebbia's workflow, slide footnotes lead back to source material, so a reviewer can check a figure against the original table or correct a reporting period before the deck goes to an investment committee. The customer story uses a hypothetical example of a banker assessing an industrial-software acquisition with roughly 2,000 files to review.

What it means for deal work

The claimed gains point to less time spent reformatting slides and checking citations, and more time spent judging a deal. Hebbia says it wants finance professionals to assess opportunities with research they can trace back to its sources. The figures are Hebbia's own test results, as described by OpenAI News, and have not been independently verified.

Kavita also described a possible future role for agents that watch information in the background and flag signals for finance professionals. She said such a system could help users start thinking about an opportunity sooner than rivals and process more chances than a person could hunt for manually. That is a forward-looking idea, not a product claim in the customer story.

Sources

  1. Hebbia sees improvements in financial research with GPT-6 AstraOpenAI News

Tags

Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.