July 12, 2026
Connection credentials in the Connect dialog, hardware-aware templates, a faster console worldwide, and near-instant async request dispatch.
Quality-of-life release — connecting to pods, picking the right template, and waiting less: for the console and for async results.
Major features
Section titled “Major features”Connection credentials in the Connect dialog
Section titled “Connection credentials in the Connect dialog”Templates that protect their web apps with a token or password (Jupyter, code editors, and similar) now surface those credentials directly in the pod’s Connect dialog — masked by default, with one-click copy. No more digging through environment variables to find the login token after deploying. Official templates have been updated to take advantage of this.
Hardware-aware templates
Section titled “Hardware-aware templates”Every template is now classified by the hardware it runs on — GPU, CPU, or both — and the console uses that to guide your choices:
- When creating a pod or submitting a task, hardware options incompatible with the selected image are locked out.
- The task image picker now shows framework runtimes only, so the list stays focused on images meant for task execution.
- New official Python 3.12 / 3.13 / 3.14 task runtimes are available, and they run on both GPU and CPU nodes.
Improvements
Section titled “Improvements”- The console loads noticeably faster, especially on high-latency connections: signing in takes fewer round-trips, API responses are compressed, the initial download is much lighter, and inference API connections are reused across requests.
- Async requests start almost immediately. Queued requests (async images, videos) are now picked up within about a second of submission, instead of waiting for the next scheduling cycle — previously up to ~30 seconds.
- The pricing page now shows a live GPU catalog and includes a cost calculator.
- Fixed an issue where a task submitted with an image incompatible with the selected hardware could fail to run while still incurring charges. Incompatible combinations are now blocked at submission.
- Requests to a model that isn’t deployed in your workspace now return a clear “model not found” error immediately, instead of being accepted and left pending.
- Fixed cases where a video generation that had actually finished could be discarded due to a mistimed status check near completion.
- Improved isolation and reliability of serverless request processing across workspaces — steadier scale-down decisions, and processing stalls are now detected and recovered automatically.