I’m Cory Torres Bishop, a product manager on Bubble’s AI team.
Today, Anthropic released Claude Sonnet 4.6 to the public — and it went live in Bubble AI at the same time. This is a milestone for us: Our team provided technical feedback that helped shape the model Anthropic is releasing today, and we’re deploying as a launch partner.
This is the first time we’ve deployed a model on the same day as its public release, and it reinforces what we’ve been building toward: the future of software development isn’t about choosing between AI speed and visual control — it’s about having both. For web and mobile.
The impact of Sonnet 4.6
The biggest jump we’re seeing is in UI generation quality. In our testing, Sonnet 4.6 is taking a more ambitious approach to layouts — it’s not playing it safe with basic structures anymore. The model consistently makes better style choices than the models we’d previously been using across the board, and the output is noticeably cleaner.
Where this really shows up is in visual design decisions. Dark mode is a good example: The new model is way more thoughtful about layering and contrast, creating interfaces that actually feel designed rather than just functional. Overall styling quality is significantly better; it’s got great “taste,” making design choices that we’d previously have had to refine manually.
On the Agent side, we’ll be rolling out Sonnet 4.6 integration very soon, and the performance improvements are substantial. Expression generation in particular (things like setting up dynamic content that pulls from your database) is running up to 4x faster. Token efficiency behind the scenes is also better, which means the whole experience of building with the Agent is faster and more efficient.
Try it out
Sonnet 4.6 is live now in Bubble AI for both web and native mobile app generation. Give it a go!
I’m trying to run a like-for-like comparison against a previous SLA app I generated with Bubble’s AI, and also against a fully wired version my agent built, just to see where this version lands in terms of actual lifecycle logic vs UI scaffolding.
But I’m currently blocked by a bug. It keeps throwing “you can only generate one app at a time,” even though I’m not generating anything else. I’ve refreshed, logged out/in, and tried again with the same result.
Is there something stuck server-side, or a way to reset the generation state?
“Great taste” is a bit too far. All of these screenshots lack common design patterns in modern-day apps (toggles, floating elements, etc.) and are extremely high in cognitive load (dense/busy). There is a LOT going on here. They seem to be from the ~2015/2016 era of design and do not really reflect a 2026 UI/UX.
Actually, compared to what I see here, the old prebuilt components were very well put together (especially the header that had mobile functionality built in).
It does seem like an improvement over the previous model, which was not capable of building a “selector.” I would be surprised if this works as expected out of the box with the expected styling and functionality. I imagine it uses a straight up color change rather than sliding across.
Selector (previous Bubble AI models didn’t ever do this):
I feel like Opus 4.6 might do a better job but of course it would be too expensive for a free tier. To be fair to the AI team, some of my complaints reflect inherent limitations in Bubble itself rather than a problem with the AI per se.
