deniz.in

Markets

Weather

Loading weather

· via OpenAI blog/news

Hebbia reports GPT-6 Astra gains in citation accuracy and deck-building for financial research

Hebbia says GPT-6 Astra beat rival models by double-digit margins on citing the right sources and following presentation briefs, in tests OpenAI published as a customer story.

Hebbia reports GPT-6 Astra gains in citation accuracy and deck-building for financial research

What was tested

Hebbia, whose Matrix platform helps finance professionals work through document-heavy research, has tested OpenAI's GPT-6 Astra on two core parts of an analyst's job: tracking down the right sources for a question and producing a client-ready presentation. According to a customer story published on OpenAI's blog on 20 November 2026, the model outperformed every alternative Hebbia had tried on both tasks.

Matrix is built for deal work of the kind that precedes an acquisition recommendation: financial statements, customer contracts and management presentations have to be synthesised, with answers tied back to the documents behind them. In a hypothetical example Hebbia describes, a banker reviewing a roughly 2,000-file data room for an industrial-software acquisition could use Matrix to examine customer concentration, contract terms and churn, pull in peer filings and industry news, and then generate a diligence summary and valuation deck in the firm's own template.

GPT-6 Astra's role in that workflow is twofold: it selects the documents or Matrix cells that support an answer, and it converts the resulting analysis into native PowerPoint slides that follow the user's structure and guidelines.

Citation accuracy

The headline numbers concern source selection. In Hebbia's tests, GPT-6 Astra scored 13.6 percentage points above the best alternative model tested on citing the required types of sources, and 99.5 percent of the excerpts it cited were relevant to the question being asked.

Hebbia evaluated a prerelease build of the model between August and September 2026, using identical financial research cases and grading criteria for each model. The company says domain experts wrote or audited those criteria against the source documents.

Traceability is central to the product. Users can open a citation to see the underlying document and the highlighted passage, and slide footnotes link back to the source material — enabling checks as narrow as verifying a customer-concentration figure against the source table, or correcting a reporting period, before a deck reaches an investment committee. Kavita Anand, a member of Hebbia's product and technical staff, said traceability matters to users "because they're basing millions of dollars of decisions on what the AI actually comes back with."

Presentation quality

The second workload was producing output a banker would actually send. Hebbia measured whether the model followed the brief — the user's structure and guidelines — as well as the overall quality of the presentation. GPT-6 Astra met 97.9 percent of the applicable presentation requirements, 14.4 percentage points ahead of the best alternative tested.

According to Hebbia, other models tended to trade off between polish and adherence to instructions: some produced well-crafted slides but ignored guidelines, while others followed the brief but looked machine-generated. Anand said GPT-6 Astra was the first model the company had seen rank first on both dimensions. The goal, as she put it, is a first draft close enough to the final output that overnight work on a deck collapses into a few minutes of review.

Agents on the horizon

Beyond the current tests, Hebbia sketched a future use for agents that monitor information in the background and surface signals for finance professionals. The appeal, as Anand framed it, is receiving a meaningful signal hours before competitors and processing more opportunities than a human could hunt manually.

Why it matters

This is one of the more concrete data points so far on how a frontier model performs against real professional workloads rather than general-purpose benchmarks. The gains Hebbia reports are not marginal: double-digit percentage-point improvements in citation coverage and format compliance translate directly into less time spent verifying sources and reformatting decks, which is where much of the labour in financial research actually sits.

Two caveats are worth keeping in mind. The results come from a customer story on OpenAI's own blog, based on Hebbia's internal testing of a prerelease build, with no independent verification and no named comparison models. The deal workflow Hebbia describes is also explicitly hypothetical. Still, the pattern is telling: in high-stakes knowledge work, the differentiators are accuracy of attribution and adherence to house formatting — not just the quality of the prose — and vendors are increasingly publishing targeted evaluations to demonstrate exactly that.

  • #openai
  • #gpt-6
  • #hebbia
  • #financial-research
  • #llm-evaluation

Related posts