ChatGPT just stopped being a text box
Issue #054 · 4 min read
ChatGPT just stopped being a text box
GPT-6 ships with charts and buttons baked in. Copilot gets a cheap, fast model. GitLab ships a lot at once.
The big story
OpenAI rolled out GPT-6 to everyone, and the bigger change isn't the model. It's the interface wrapped around it.
ChatGPT answers now come back as pictures, charts, and buttons instead of paragraphs — that's The Verge. OpenAI calls it "Intelligent UI": faster responses with visuals and interactive elements you click through instead of read, according to OpenAI's own announcement.
Ask ChatGPT something and instead of three paragraphs you might get a chart, a comparison table, and a row of buttons to narrow your question.
TechCrunch frames it as ChatGPT becoming "a lot more visual." That's the polite version. The blunter version: OpenAI is betting that a chat box was never the right shape for most questions, and it's reshaping the product to look less like a search bar and more like an app.
For product people, this is the thing to actually sit with. Every product that wraps a prompt in a text field — which is most of what got shipped in the last two years — now looks like it's competing against the default experience, not building on top of it.
If your product's entire value proposition was "we put a nicer UI on top of a model's text output," OpenAI just moved that goalpost for free.
What shipped
GitHub Copilot added Claude Haiku 5.5, Anthropic's lightweight model, built for fast, high-volume tasks like subagents and quick edits — GitHub's changelog. It's also live on Vercel's AI Gateway the same day.
Two platforms adding the same cheap, fast model within hours of each other is a signal worth reading: the race isn't only for the smartest model anymore, it's for the cheapest one good enough to run constantly in the background.
If you're pricing an AI feature on token cost, this is a lever worth pulling before your next planning cycle.
GitHub also shipped a purpose-built model just for catching leaked secrets in code, designed to keep pace with how much code AI agents now write — per the changelog. Pair that with local sandboxing for Copilot going GA across the CLI, desktop app, and VS Code — also GitHub — and the pattern is clear.
As agents write more code unsupervised, GitHub is spending real effort on the guardrails around that code, not just the code itself. Worth a look if your team is letting agents commit without a human reviewing every line.
GitLab bundled over a dozen announcements into one release they're calling Transcend, covering agent orchestration, dependency security, and build governance — see GitLab's roundup. The one worth your actual attention is Dependency Firewall, which blocks risky open-source packages before a build runs.
GitLab says its researchers found five malicious PyPI packages in June, four of them typosquats of Flask, Requests, and NumPy, sitting quietly in install scripts waiting to steal CI/CD credentials. Their stated worry isn't just human developers anymore — it's that coding agents now add dependencies on their own, often unreviewed, which means nobody is checking what's sneaking in. If your engineers are running agents with write access to package.json or requirements.txt, that's the gap this closes.
What I'd actually do this week
Open ChatGPT and actually look at how an answer renders now, not just what it says. If your product shows users a wall of AI-generated text, ask whether that's a deliberate choice or just the easiest thing to build.
Check which model is running your highest-volume, lowest-stakes AI calls — autocomplete, quick edits, simple classification. If it's not something Haiku-class, you're probably paying for intelligence you don't need there.
If anyone on your team lets an agent add dependencies without review, ask them to read GitLab's typosquat example before Friday. Five minutes, real cost if ignored.
Reply and tell me what Intelligent UI looks like once you've actually poked at it — I'll compare notes with other readers in a future issue.