Every model now carries its own effort level — and the Claude 5 models decide for themselves when a step deserves more thinking. Alongside that: the assistant really watches a video you attach — including the recording of an old application you hand over when you create a project — and a running turn no longer blocks your other projects.
Every demo runs on its own — this is how it looks in the workspace.
Every model in the picker carries its own control: Low, Medium, High — side by side, without switching model or closing the menu. Claude Fable 5, Opus 5.5 and Sonnet 5 additionally offer Adaptive and start there: the model decides per step.
An old application whose source code is long gone: record it running and attach the recording when you create the project. The assistant watches it frame by frame — screens, fields, steps — and builds the first version from that.
A screen recording could be uploaded and shipped — but not watched. The assistant now sees it as timestamped frames: a coarse pass to find the moment, then a closer look at the seconds around it.
A running turn blocked every other project in your account — half an hour, on a long turn. The lock guarded nothing the database was not already guarding, and it is gone.
A publish can stop twice for a decision — open endpoints, security findings. Both dialogs used to stand on screen at the same time. Now one question is asked, and the second follows if it still applies.
Improvements across the platform — from chat and attachments to billing and published sites.