Pick a Deep Intelligence model from the prompt bar dropdown
The intelligence control in the prompt bar now opens a compact dropdown where Deep Intelligence sits under Automatic, with its models listed so you can pin the one you want.
Changing how much thinking Chief puts into an answer used to pull you out of the prompt bar into a large separate surface, and the strongest tier looked like something off to the side rather than a choice Chief could make for you. Picking a specific model for that tier was close to unreachable.
Compact intelligence dropdown
The intelligence chip in the prompt bar now opens a small dropdown anchored to it, marked with a down arrow. The tiers appear as single-line rows and one click applies your pick and closes the dropdown. The longer explanations and the fuller view are still there from their other entry points.
Deep Intelligence under Automatic
Fast, Expert and Deep Intelligence now hang off Automatic as siblings, because Automatic can pick any of the three based on what you ask. The rules connecting them are drawn in a neutral gray that stays visible in dark mode.
Model choice inside Deep Intelligence
Switching to Deep Intelligence reveals the models available on your plan instead of closing the panel before you get the chance. Each model shows how many credits it draws, and a checkmark marks the one in use. Power User plans see Claude Fable 5 alongside Claude Opus 5; other paid plans get Claude Opus 5. You can also save your tier, vendor and model as the setting for new chats.
Updated model lineup
Deep Intelligence now runs Claude Opus 5, GPT-5.5 and Gemini 3.5 Flash, and the AI Provider list names the same models the tier actually runs, per vendor. Expert runs Claude Sonnet 5, GPT-5.4 Mini and Gemini 3.1 Pro. Agents work in every tier now, so mentioning an agent no longer forces you into Deep Intelligence: it runs at whatever level the chat is set to.
Retry on another model after a refusal
When a vendor's safety classifier declined a request, the turn used to come back as an empty response after burning retries. You now get a card that names what happened, with technical details you can expand and a button that resends the turn on a different model, named up front, such as "Try with GPT-5.5". Once all three vendors have declined, the button disappears and Try Again remains. Turns that spend their whole budget reasoning now show that reasoning instead of arriving blank.
Open the intelligence chip in the prompt bar to see the tiers nested and pick your Deep Intelligence model.