# GPT-6 Astra beat the next-best model by 14.4 points in Hebbia's banker deck tests

By MarketScale Newsroom · Published 2026-11-22 · Software & Technology on MarketScale
Canonical: https://www.marketscale.com/industries/software-and-technology/gpt-6-astra-beat-the-next-best-model-by-144-points-in-hebbias-banker-deck-tests

> Hebbia's prerelease tests put GPT-6 Astra first on sourcing and slides. Its credit study shows where a far cheaper model gets close.

## Key points

- Hebbia's own numbers suggest a two-model shop: GPT-6 Astra led on citations and client-ready slides, while a post-trained open model came within three points of Astra on credit tasks at roughly one-seventh the run cost.
- Score sourcing on two separate measures, coverage of required source types and relevance of each cited excerpt, because a model can pass one and fail the other.
- In Hebbia and AfterQuery’s credit study, Nemotron’s pre-training failure was missing qualifiers or exceptions even after finding the right clause—making completeness a key test alongside retrieval.

Every number in a diligence deck has two readers to get past. One is the partner who checks it against the data room. The other is the client, who will notice if the slides look machine-made. Hebbia, whose Matrix platform links research answers to the documents behind them, says a prerelease build of OpenAI's GPT-6 Astra beat every alternative it tested on both counts.

The results come from tests Hebbia ran in August and September 2026, which OpenAI published as a customer story on November 20. The headline gap is on slides: Astra led the next-best model Hebbia tried by 14.4 percentage points on meeting presentation requirements. That's a strong showing on the part of the job that, by Hebbia's account, produces the overnighters.

Hebbia has also published a second set of numbers. It ran a credit study with AfterQuery, and that one shows a far cheaper model getting close to Astra. Together, the two point anyone buying AI research tools for a bank, a fund or a corporate development team toward picking models task by task.

## Inside Hebbia's Matrix test

Hebbia tested two steps. First, the model picks the documents or Matrix cells that support an answer. Then it turns that analysis into native PowerPoint slides that follow the user's structure and guidelines. Hebbia illustrated the flow with a hypothetical case: a banker assessing a client's acquisition of an industrial-software company, with about 2,000 files in the data room. The banker digs into customer concentration, contract terms and churn, then needs a diligence summary and valuation deck in the firm's template.

The method is the part worth copying. Every model ran through the same financial research cases and was scored against the same grading criteria. Domain experts wrote those criteria or audited them against the source documents.

Block Field

The two sourcing numbers catch different failures. The 13.6-point lead measures coverage, meaning whether the model brought in the kinds of sources a question requires. The 99.5% figure measures precision, meaning whether each excerpt it cited was relevant. A model could cite every required source type and still pad its answers with irrelevant passages. It could also cite cleanly from too narrow a set of documents. Hebbia reports that Astra did well on both.

The slide results fit the same pattern. Kavita Anand, a member of Hebbia's technical product staff, said earlier models tended to be strong either on slide craft or on following the brief and guidelines, and rarely on both. Astra was the first model Hebbia had seen rank first on both. For a firm with strict brand templates, that combination could decide whether an AI draft gets edited or rebuilt from scratch.

## Why the citation trail matters most

Anand's case for traceability is about money. Hebbia's users base decisions worth millions of dollars on what the AI returns, she said, so citations do two jobs: they let a reviewer trace a claim, and they give confidence that the model found the right answer. In a data room holding thousands of documents, she added, even a one-point gain adds up.

In Matrix, a reviewer can open a citation to see the underlying document with the passage highlighted, and slide footnotes lead back to the same material. Hebbia's example of a review is specific: check a customer-concentration figure against the source table and correct the reporting period before the deck goes to the investment committee. The human check stays in the process. What the model changes is how much of the reviewer's time goes to verifying and reformatting, and how much is left for judging the deal.

Anand expects a better first draft to turn all-night deck sessions into a few minutes of review. That is a projection, not a measured result. Trendoviz notes that the story names none of the alternative models and gives no full per-model scores, so the leads can't be rebuilt independently, and that the scores come from a prerelease build.

For a deal team piloting Matrix or a similar tool, those gaps point to the number worth tracking in-house: time from first draft to an approved deck, measured on the team's own deals. Benchmark leads belong in the vendor conversation. Review hours belong in the business case.

## Where a cheaper model gets close

Hebbia's other study tested a different kind of work. AfterQuery and Hebbia post-trained NVIDIA's open Nemotron 3.5 Super VL model on expert-written credit tasks built around syndicated loan agreements and their amendments. On a held-out set of 100 tasks, the share fully solved rose from 47% to 74%. GPT-6 Astra solved 77% with no extra training. AfterQuery puts the cost of the full 100-task run at $24.72 for Astra and an estimated $3.63 for the post-trained Nemotron, roughly one-seventh.

Block Field

The method matters for reading the chart. Each model ran each task once, and a judge model graded every answer against an expert-written rubric. Between three and eight tasks per model were dropped because the judge recorded no score. Fable 5.1 was the most accurate model at 85%, with a run cost of $49.66, about 14 times the post-trained Nemotron's. GPT 6 Sol scored 63% at $5.35. Both Nemotron cost figures are estimates based on token counts and OpenRouter list pricing.

The failure the training fixed is instructive. Before post-training, Nemotron had the clause that decides the answer on screen in 99% of runs, yet its answers still left out a qualifier, an exception or half of a two-part question. After training, the number of grading criteria failed for missing content fell from 58 to 13, and section citations per answer rose from one to 10. The model already found the right clause. What it learned was to put the proviso in its answer, and in credit work the proviso is often the answer.

Scale explains why Hebbia cares. Its customers manage more than $30 trillion in assets, according to AfterQuery, and one loan can sit under a stack of amendments while a portfolio can mean thousands of documents. Adithya Ramanathan, Hebbia's head of applied research, said the results show credit expertise can be applied to tens of millions of documents at a practical cost. AfterQuery's suggested setup is a post-trained open model for high-volume, well-scoped work, with edge cases sent to a frontier model for extra review.

Hebbia's own numbers suggest a two-model shop: GPT-6 Astra led its banker tests on citations and slides, while a post-trained open model came within three points of Astra on credit tasks at roughly one-seventh the run cost.

There are limits to that reading. The deck tests and the credit tasks measure different work, so Astra's 97.9% and 77% can't be compared. The $3.63 also covers only running the tasks. The expert task writing, Hebbia's credit-workflow taxonomy and the post-training run on 32 H200 GPUs all sit outside it. A firm without former credit investors on staff to write and check tasks would be starting further back.

## Running a model bake-off the Hebbia way

Both Hebbia studies share a structure a buyer can borrow: a fixed set of cases from real workflows, grading criteria written or checked by people who do the work, and every model run side by side. That's a test a research technology lead can run on internal deals before signing a contract, whichever vendor is quoting the win.

Block Field

Three steps follow from the published methods. Score source coverage and citation relevance separately, as Hebbia did. Record cost per task next to accuracy, the way the credit study's chart plots both. And grade completeness directly, including whether an answer kept the qualifier or exception, since that was where the base model in the credit study kept falling short.

The next thing to watch is agents. Anand described a future role for agents that monitor information in the background and flag signals to finance professionals hours before competitors see them. Hebbia presents this as a direction, not a shipped feature. An agent like that would run constantly across huge document volumes, which is exactly the job where Hebbia's credit study found cost per document mattering as much as the top score.

## Sources

- [Hebbia sees improvements in financial research with GPT-6 Astra](https://openai.com/business/customer-stories/hebbia/)
- [Post-Training NVIDIA Nemotron 3.5 Super VL for Finance - AfterQuery](https://www.afterquery.com/blog/post-training-nemotron-on-credit-workflows-in-partnership-with-hebbia)
- [Hebbia tests GPT-6 Astra on financial research - Trendoviz](https://trendoviz.com/trends/hebbia-tests-gpt-6-astra-on-financial-research)

Tags: Hebbia, GPT-6 Astra, OpenAI, AfterQuery, NVIDIA Nemotron, AI model evaluation

---
Source: MarketScale, https://www.marketscale.com/industries/software-and-technology/gpt-6-astra-beat-the-next-best-model-by-144-points-in-hebbias-banker-deck-tests. Published for AI indexing and citation; cite the canonical URL. Site guide for agents: https://www.marketscale.com/llms.txt
