GPT-6 Astra models · connector inheritance · steadier sessions

You can now pick the GPT-6 Astra lineup when starting a Task or switching models mid-conversation.
Claude Fable 5.1 cache pricing is now reported accurately, so credit usage matches what you actually spend.
Claude models are routed to whichever provider is healthy at the time, which reduces "model unavailable" interruptions.
When a provider rejects a request for billing reasons, the chat now shows a clear terminal error instead of silently retrying.

Connector settings now cascade from bot to project to task, so a task opened inside a bot-owned project inherits the right connectors automatically.
Moving a task between projects immediately refreshes its connector owner, so the banner no longer shows the previous project or bot.
A partially enabled connector set now displays its real state instead of appearing fully off.
Offline connectors are no longer offered when you create a scheduled task.
Toggling a connector without permission now shows a plain explanation instead of a generic error.

Recently deleted spaces are now listed for recovery no matter your role in the space you are currently viewing.
Dragging a space to the very top or bottom of the list now always lands where you dropped it.

Staying signed in across a page refresh is more reliable: the session is no longer dropped while your profile is still loading.
Opening the app is faster thanks to fewer repeated account lookups and better database connection handling.

kollab model run can now run any chat model from the catalog directly from your terminal.
The CLI refuses to reuse credentials saved for a different Kollab environment, which prevents accidental cross-environment calls.
Standalone API usage and cost reporting is more accurate, and a response with no usage is no longer settled as free.