Auto tries your own subscription first; Desktop retires

When you pick Auto, we now try the models your own AI subscription serves before anything else. If you connected ChatGPT and your workspace shares a Claude plan that serves Opus, a conversation on Auto runs on your ChatGPT first and reaches the shared Opus only after it. Model lists now read by capability, most capable first within each family, so Claude reads Fable 5.2, Opus 5.5, Sonnet 5.5, then Haiku 5.5. And General Input Desktop is retired: your subscriptions run in our cloud through the Token Tunnel, and the desktop app no longer connects.

How it works

• Yours, then shared, then hosted. Auto first tries the models your own accounts serve right now, in the tier's usual order, then the ones only a subscription shared with your workspace serves, then General Input's hosted models. This holds in every workspace.
• No step down for the work. Fast, Balanced, and Max each try only the models chosen for that tier, so preferring yours changes which account runs the work, not whether the model suits it.
• A limit falls through. When your account reaches its limit, the work moves to a shared copy of the same model, even mid-run.
• Desktop users sign in once in Settings. An installed desktop app now stops on its own. Sign in to the same accounts under Models and your workflows keep their models. A workflow set to a model only the app served stays on: until you sign in, it follows its If offline setting, skipping the run or using a hosted model, and never turns itself off.
• Lists by capability. The model picker on web and phone, and the Subscriptions and Models pages in Settings, list each family from most to least capable, not by price.