Extended reasoning
Give the model room to think before it answers, and decide how much.
The Think button turns on extended reasoning for the next message. The model works through the problem before committing to an answer, which helps on multi-step problems and hurts on simple ones by making them slower.
Turning it on
Tap Think above the input. A budget selector appears next to it.
| Budget | Reasonable for |
|---|---|
| 1K | A nudge. Barely more than normal. |
| 5K | Short multi-step problems. |
| 10K | The default, and the right answer most of the time. |
| 25K | Involved analysis, tricky debugging. |
| 50K / 100K | Genuinely hard problems where you would rather wait. |
The budget is a ceiling on reasoning tokens, not a target. A question that does not need 100K will not consume it.
When it is worth it
Reach for Think when the question has steps that depend on each other — working through a proof, tracing a bug, weighing several options against constraints, planning something with real interdependencies.
Leave it off for recall, formatting, summarizing and rewriting. Reasoning does not make those better; it just makes them slower.
Reasoning is not the answer
The model’s working is not shown to you as the reply. AODex deliberately separates the two: what you get back is the conclusion.
If you see The assistant reasoned about this but didn’t produce a final answer, the model spent its budget thinking and stopped before writing the conclusion. Send the message again — usually with a larger budget, or with the question broken into smaller pieces.
Availability
A persona can switch Think off, in which case the button is not shown. Not every model supports extended reasoning either; on those, the button has no effect worth paying for. See Choose a model.