The data viz problem nobody wanted to admit
Data visualization has been LLMs' awkward gap for years. Models that could write essays, debug code, and reason through math would confidently spit out broken Matplotlib spaghetti or aesthetically criminal bar charts. You'd get technically correct output that no analyst would ever show their CEO.
Hex is betting GPT-6 Astra changes that. According to Caitlin Colgrove, Hex's co-founder and CTO, the new model doesn't just generate charts—it builds "complex, interactive, compelling, and beautiful data artifacts" that people are "actually proud to share with the rest of the organization."
That's a big claim. Let's dig into what Hex is seeing that makes them think viz is finally a solved capability.
Geospatial transforms and aesthetic judgment
The hard part of data visualization isn't drawing shapes. It's the unglamorous pipeline work: reshaping dataframes, handling nulls, projecting coordinates, picking color scales that don't lie. Caitlin points out that GPT-6 Astra can "go into the underlying libraries, do all of the crazy transformations that geospatial visualizations require, and build out something that's not just functional but aesthetically pleasing."
Geospatial is the stress test here. If you've ever wrestled with GeoPandas coordinate reference systems or tried to get Folium to render a clean choropleth, you know there's a huge gap between "technically runs" and "looks good in a stakeholder deck." The fact that Astra can navigate the underlying library APIs and make aesthetic calls suggests meaningful multimodal reasoning about visual output.
Hex's example workflow: ask how sales channels are performing, which are strong, which should be discontinued. The system generates written findings, trend lines, comparisons, and interactive dashboards with geographic breakdowns. That's not a single visualization—it's a composed analytical artifact with multiple views coordinated around a business question.
Analytical judgment: the harder capability
Here's where it gets interesting. Hex isn't just using GPT-6 Astra to write plotting code. They're using it to interrogate the answer.
Caitlin describes this as "analytical judgment"—the model asking itself: Does this number make sense? Does the analysis actually answer the user's question? Does it reflect the business objective? She says Astra "investigates more thoroughly, uncovers more of the data's pitfalls and nuances, and delivers a more comprehensive analysis than earlier models."
This is a different kind of capability than "generate a bar chart." It's closer to what a senior analyst does: sanity-checking results, considering edge cases, thinking about whether the framing matches what the stakeholder actually cares about. If the model can catch when technically correct SQL produces a misleading metric (because of how nulls propagate, or a weird spike in one region, or a mismatch between fiscal and calendar quarters), that's huge.
It also means Hex is running Astra in a loop: generate analysis, critique analysis, refine. That's architecturally different from one-shot code generation. It suggests Hex has found a sweet spot where the model's reasoning is reliable enough to self-correct without spiraling into overthinking or hedging.
Making data work shareable
Caitlin frames the core problem clearly: "A really important part of data work is not just doing the analysis, but also communicating it." The output has to be something colleagues can engage with—comparing results, exploring details, following changes over time.
This is where interactive dashboards earn their keep. A static PNG of a trend line doesn't let your VP of Sales drill into regional performance. A well-composed Hex report with filters, breakdowns, and linked views does.
The pitch is that GPT-6 Astra makes it trivial to go from question to that polished, explorable artifact. No more "let me spend two hours in Plotly to make this not embarrassing." The model handles the transforms, the layout, the interactivity config, and the aesthetic choices that make the difference between a draft and something you'd put in front of executives.
What this tells us about GPT-6 Astra's capabilities
Hex's use case is a useful lens on where GPT-6 Astra might have genuinely improved over earlier models:
- Deep library knowledge: navigating geospatial library APIs and data transform chains requires detailed awareness of how specific packages work, not just generic Python syntax.
- Compositional reasoning: building a multi-view dashboard means coordinating several visualizations around a shared analytical narrative.
- Aesthetic judgment: making choices about color, layout, and clarity that go beyond "does the code run."
- Self-critique loops: interrogating whether an answer is correct, sensible, and aligned with the user's intent.
That last one is particularly interesting. It suggests Astra can hold a representation of "what would a good analysis look like here" and compare its output against that standard. If that's robust, it's a big step toward agentic data work that doesn't require constant human supervision.
The broader data agent picture
Hex positions itself as an "agentic data platform," and this integration shows what that might actually mean in practice. Not just "chatbot that writes SQL" but a system that can:
- Understand a business question
- Query and transform data
- Generate visual artifacts
- Critique and refine its own work
- Produce something ready to share
If GPT-6 Astra is genuinely better at steps 3-5, that closes a major gap in the data agent workflow. Previous models could get you partway there, but you'd still need a human to clean up the viz and sanity-check the conclusions.
Caitlin's comment that Hex wants to make "everyone a data person" is the ultimate test. Can non-analysts ask good questions and get trustworthy, presentable answers? That requires not just technical correctness but judgment, clarity, and polish.
What we still don't know
Hex's case study is encouraging, but it leaves open questions:
- Failure modes: How often does Astra generate a beautiful dashboard that answers the wrong question? What does the error UX look like?
- Customization boundaries: Can users override aesthetic choices or library defaults when the model's judgment misses the mark?
- Cost and latency: Running multiple critique loops with a frontier model isn't cheap. What's the user experience when you're waiting for Astra to refine its third draft of a geographic breakdown?
Those aren't criticisms—Hex is clearly seeing enough value to build on Astra. But they're the operational realities that determine whether this works at scale across an organization.
Why this matters
Data visualization has been a stubborn weak spot for LLMs because it requires taste, domain knowledge, and compositional reasoning all at once. If GPT-6 Astra is genuinely closing that gap—handling complex transforms, making aesthetic calls, and exercising analytical judgment—it's a meaningful step toward systems that can do end-to-end data work without constant human intervention.
Hex's integration suggests we're moving past "models that write code" toward "models that produce artifacts humans want to use." That's a higher bar, and it's the one that actually matters for making AI useful in business contexts.
We'll see how it holds up in the wild. But if Caitlin's right that analysts are now proud to share what Hex generates, that's a signal worth paying attention to.