Auto runs on your Token Tunnel subscriptions first

Auto picks the model for each request. It now considers the models you've connected through Token Tunnel and prefers them: when a suitable one is online, Auto runs there, on the subscription you already pay for, instead of on a hosted model billed in credits. A Default to local models switch in the Auto menu controls it; it's on by default once you have a connected model.

How it works

• Tiers. Auto classifies a request as an answer, an execution, or a build, mapped to Fast, Balanced, or Max. Anything it can't classify goes to Max.
• A ladder per tier. Each tier has an ordered list of Token Tunnel models, and every tier includes all four providers (Claude Code, Codex, Grok, Gemini), so a single subscription is enough.
• Never a blocker. If the switch is off, nothing in the ladder is online, or a lookup fails, the tier runs its normal hosted fallbacks.
• Every subscription before the cloud. When one subscription hits its limit mid-run, the run moves to the next online one, and falls back to a hosted model only after all of them are exhausted.

Also shipped

• Smaller models that couldn't reliably finish agent work became opt-in, and anyone defaulted to one moved to Auto.
• Gemini 3.7 Flash replaced Gemini 3.6 Flash, and Settings, Models now shows who hosts each model and where.