Image: METAL
Summary
- OpenAI published a startup case study on September 16 describing how Hex uses GPT-6 Astra.
- Caitlin Colgrove, co-founder and chief technology officer of Hex, said models have struggled with data visualization for a long time.
- Hex has Astra ask whether a number makes sense, whether the analysis answers the question, and whether it reflects the business objective.
The hardest part of data analysis is not the code but communicating the result to someone else. In a case study OpenAI posted on September 16, Caitlin Colgrove, co-founder and chief technology officer of the agentic data platform Hex, said that a really important part of data work is not just doing the analysis but also communicating it. In the same piece she noted that models have struggled with data visualization for a really long time, and said GPT-6 Astra changed that.
According to the announcement, Hex puts Astra inside its data agents to turn answers into interactive visualizations. OpenAI filed the piece as a startup case study and listed the region as North America, the industry as technology, and the product used as the API. The page carries a 48-second video and a one-minute-45-second audio version.
Colgrove described the difference in the output this way. With GPT-6 Astra, she said, you can build much more complex, interactive, compelling and beautiful data artifacts that not only communicate your analysis clearly but that you are actually proud to share with the rest of the organization. It is a sentence that measures the value of a visualization by shareability rather than accuracy.
The technical account is more specific. Producing a usable visualization requires both engineering work and choices about presentation, and Colgrove said Astra goes into the underlying libraries, performs all of the transformations geospatial visualization demands, and builds something that is not just functional but aesthetically pleasing inside the data artifact. Layering values onto a map combines coordinate conversion with boundary data joins, and it is the kind of work analysts abandon most often.
The example OpenAI gave involves sales channels. A Hex user asks how channels have performed over time, which are doing best, and which should be discontinued, and Hex returns written findings alongside trend lines and comparisons. Interactive dashboards let users explore key measures, rankings and geographic breakdowns directly.
The second use is the one engineers should watch. Hex uses Astra not only to produce an answer but to interrogate the one it produced. Does this number make sense, does this analysis answer what the user asked, does it reflect the business objective. Technically correct code frequently returns an answer that misses the point.
Colgrove calls the capability analytical judgment. Astra, she said, is much better at exercising what they call analytical judgment, which is not just getting to an answer but understanding whether that answer is actually what the user is asking for and what the business needs. OpenAI wrote that Astra investigates more thoroughly and uncovers more of the data's pitfalls and nuances than earlier models.
Why that verification loop sits at the center of the product is recorded in Hex's own history. Introducing its Notebook Agent on August 27, 2025, the company wrote that a SQL query can execute successfully and still produce disastrously wrong results. In the same post it argued that data analysis is not a verifiable task and demands taste, intuition and business judgment.
The mechanisms the company added then target the same problem: mentions that point the agent at specific cells and tables, versions saved automatically for every thread, and a diff view that surfaces or rolls back agent-written code cell by cell. The Notebook Agent opened in public beta to editor-level users and above on paid plans, and at announcement the model driving it was Claude Sonnet 4. Ian M., head of analytics engineering and data science at Ramp, said in that post that exploratory analysis sped up when juggling multiple datasets.
The case for Astra in this role rests on figures OpenAI published with the model. In that announcement, which METAL reviewed, Astra scored 59.3 percent on Agents' Last Exam, which tests professional tasks in real software, against 55.5 percent for Claude Opus 5 and 53.6 percent for GPT-5.6 Sol, while using roughly 65 percent fewer output tokens than Opus 5 at the same setting. In OSWorld 2.0 latency simulations it reached 72.6 percent at about 40 minutes per task, against 65.7 percent at about 75 minutes for Sol, cutting time by roughly 47 percent.
Figures on visual judgment appear alongside. On BenchCAD, which asks models to reconstruct three-dimensional objects from multi-view renders as CAD code, Astra reached a 95.9 percent geometric-overlap score with tools, above 83.3 percent for Sol and 84.3 percent for Claude Fable 5.1. Drawing a picture and judging whether the picture is right improved inside the same model.
This is not the first case of its kind. METAL has reported that the game studio Playco cut manual fixes in prototyping by 50 percent with Astra. OpenAI links two more pieces in the same format below the Hex story: a September 3 post on Legora reviewing 41 documents in minutes and the Playco piece.
The conditions on the model deserve equal attention. METAL has reported that Astra shipped with both improved alignment and a critical threshold rating in cybersecurity. According to the announcement, Astra is rolling out to ChatGPT Plus, Pro, Business and Enterprise users and is available through the OpenAI API, Microsoft Azure and AWS Bedrock. Enterprise administrators must enable it for their workspace, and access is off by default at launch.
Hex's stated goal is making everyone a data person. OpenAI wrote that with Astra, Hex users can access, understand and communicate data more easily, bringing the company closer to that goal. Competition among analysis tools is shifting from writing good queries to explaining results and checking their own work.
Comments