What Deep Think is and when to use it

Some questions have a right answer that a quick reply usually gets wrong. A multi-step calculation. A plan with real tradeoffs. A piece of analysis where the first thing that comes to mind is close, but not quite right.

Deep Think is for those. It's a Pro feature that tells the model to slow down and work the problem before it answers, instead of replying with its first pass.

What Deep Think does

Normally the model reads your message and writes back in one go. With Deep Think on, it runs the model you're on with extended reasoning: it plans the problem, works through it, and checks its own answer before showing it to you.

You can watch that happen. Instead of a blank pause, the reasoning streams as it goes, so you see the model reason through the steps rather than wait on a spinner.

Two things follow from that. It takes longer than a normal reply. And it uses more of your plan's monthly credit, because more thinking means more work. Both are fine when the problem is hard — and both are why it isn't the default for everything.

Auto already does a lighter version of this on its own. It thinks longer when it senses a problem is genuinely hard. Deep Think is the deliberate version: you turn it on when you already know the problem is hard and want the model to take its time.

When it helps

Deep Think earns its keep when getting to the answer takes more than one step:

• Multi-step reasoning. Math and logic problems, or anything where step two depends on getting step one right.
• Tricky analysis. Weighing options against each other, reviewing a contract or a piece of code for what's wrong, working out why something isn't adding up.
• Planning. Breaking a goal into an ordered set of steps, where a rushed answer skips something that matters.
• Problems where the obvious answer is often wrong. If you'd double-check the reply yourself before trusting it, Deep Think is doing that checking for you.

When it's overkill

For most messages, it isn't worth it. A quick reply is already right, and the extra time and credit buy you nothing.

Leave Deep Think off for:

• Simple questions — a fact, a date, a definition, a quick conversion.
• Short rewrites, translations, and tidying up text.
• Anything a fast model answers well the first time.

Turning it on for "what's the capital of France" just makes you wait longer for the same answer. Keep it off by default and reach for it when a problem actually needs the care.

Turning it on

Deep Think is part of Pro and Max. On the Free and Standard plans it isn't available.

You switch it on per message. You'll find it in the message bar's menu, labeled Deep Think — turn it on for the hard question, then turn it back off for the next one. You don't have to pick a model: on Auto, Deep Think runs the strongest reasoning model your plan includes. If you've pinned a specific reasoning model, it runs that one.

Plans and what's in each change over time; the current details are on the pricing page.

In short

Deep Think trades a little time and credit for a more careful answer. Use it when a problem has real steps to work through, and leave it off when a quick reply already does the job. On Auto it picks the reasoning model for you, so the only real decision is whether the question is hard enough to be worth the wait.