AI · Nov 9, 2026
OpenAI’s Control Room puts a fail-closed kill switch inside the inference engineFor Hebbia’s bankers, the citation is the product and the slide is the test
· Updated · Signal Desk newsroom
Analysis of reporting by OpenAI. How we work

Hebbia’s prerelease tests credit GPT-6 Astra with better source coverage and on-brand decks. The numbers are vendor-friendly, but they show what finance buyers now reward: traceability over eloquence.
Key takeaways
- Hebbia tested a prerelease GPT-6 Astra build in August and September 2026 on financial research cases graded by domain experts.
- Astra scored 13.6 percentage points above the best alternative tested on citing the required source types, with 99.5% of cited excerpts relevant.
- It met 97.9% of applicable presentation requirements, 14.4 points ahead of the best alternative on that measure.
- Kavita Anand of Hebbia’s product technical staff says earlier models excelled at either craft or instruction-following, not both.
- The story is an OpenAI customer case study; competitors are not named and full methods are not published.
What Hebbia tested
Hebbia sells Matrix, a research tool that links answers to the documents behind them for bankers and investors working through data rooms. In an OpenAI customer story, the company says it tested a prerelease build of GPT-6 Astra in August and September on two jobs: finding the sources a question actually requires, and turning that research into a presentation that follows a banker’s brief. Each model faced the same cases and grading criteria, which domain experts wrote or audited against the source documents.
The results favor Astra on both counts. It scored 13.6 percentage points above the best alternative on citing the required types of sources, and 99.5% of the excerpts it cited were relevant to the question. On decks, it met 97.9% of applicable presentation requirements, 14.4 points ahead of the next best model. Kavita Anand, a member of Hebbia’s technical staff on the product side, says earlier models tended to shine at either visual craft or following guidelines, and Astra was the first to rank first on both.
Read the fine print before the headline
This is a vendor story published by the model vendor, and it should be weighed that way. The alternative models are unnamed. The metrics are Hebbia’s own. The example workflow, a banker reviewing roughly 2,000 files for an industrial-software acquisition, is explicitly hypothetical. None of that makes the numbers false, but it means they describe one company’s evaluation, not an industry benchmark.
What is more useful is what Hebbia chose to measure. Not fluency, not reasoning puzzles, but whether the model pulls the right kinds of documents and whether every cited excerpt actually supports the answer. Anand’s point is that users are making decisions worth millions on what comes back, so the citation is what earns trust. In a data room with thousands of documents, she argues, even a one-point gain in coverage matters.
The deck problem is real
The second half of the test is less glamorous and probably more important to adoption. Bankers do not want a slide that looks machine-made; they want their firm’s template, their partner’s structure and nothing a client would flag. Hebbia’s workflow has Astra build native PowerPoint slides with footnotes that lead back to the source passage, so a reviewer can check a customer-concentration figure against the original table before the deck goes to committee.
That design choice is the right one. It keeps a human checking numbers at the moment they matter most, rather than trusting a polished artifact. It also fits a pattern across finance. OpenAI’s own ChatGPT for Financial Services pitched hosted data with granular citations, and Jump Trading’s agent setup put a human approval gate in front of anything that advances. Across very different firms, the winning formula is the same: traceable output and a person who signs off.
The quieter ambition
Anand also sketches a future in which agents monitor information in the background and surface signals for finance professionals, getting them to an opportunity hours before competitors. That is where the stakes rise. A tool that drafts a deck from documents a banker chose is bounded. A tool that decides which signals deserve attention starts shaping where capital looks. The traceability discipline Hebbia emphasizes here will matter even more there, and so will the question of what happens when every firm’s agents flag the same signal at the same moment.
Who feels it
- Investment bankers and analysts
- Less time reformatting decks and chasing sources, more time on the deal, if the citations hold up under review.
- Junior staff
- The grinding first-draft work that trains new analysts is exactly what tools like Matrix aim to absorb.
- Rival model vendors
- Hebbia’s criteria show finance buyers grading on citation relevance and template fidelity, not general benchmarks.
- Compliance teams
- Source-linked footnotes make review easier, but responsibility for errors still sits with the firm.
What to watch
- Whether Hebbia publishes its evaluation methods or names the models it compared against.
- How the production version of GPT-6 Astra performs versus the prerelease build Hebbia tested.
- Whether background-monitoring agents ship in Matrix and how they are kept auditable.
FAQ
What did Hebbia find in testing GPT-6 Astra?
In its tests, Astra scored 13.6 points above the best alternative on citing required source types, had 99.5% relevant citations and met 97.9% of presentation requirements, 14.4 points ahead of the next best model.
Which models did Astra beat?
The case study does not name the alternatives Hebbia tested.
Is this an independent evaluation?
No. It is Hebbia’s internal evaluation of a prerelease build, published as an OpenAI customer story.
Companies: OpenAI
More on OpenAI
AI · Nov 11, 2026
OpenAI’s $2 million for children’s AI safety research matters for one promise: publishing bad newsAI · Nov 17, 2026
OpenAI says its models now shave area off production chips, including AMD’s next flagshipsAI · Nov 3, 2026
OpenAI’s GPT-5.6 Cyber reaches AWS GovCloud through Daybreak Red, with two approval gatesAI · Nov 5, 2026
Perplexity Teammates bets that always-on work agents are worth $200 a month