Every AI app now has a dial for how hard it thinks before it answers, and every company gave it different names. This page explains what the dial does, when to turn it up, and what the names mean in Claude, ChatGPT and Gemini.
Time to work the problem out before answering.
Without thinking, a model answers straight away, the way you answer "what is the capital of France". With thinking turned up, it first works in private: breaks the task into steps, tries an approach, checks it, sometimes starts over. Then it writes the answer.
Start in the middle. Turn it up only when the answer was wrong or shallow, and turn it back down afterwards.
Seconds. Uses the least allowance.
Right for most requests.
Up to a minute or two. Use when Normal slipped up.
Minutes. Burns allowance fast; often needs a higher plan.
The rows are a rough match, not an exact one: each company sets its levels differently. Names checked on 2 October 2026; they change several times a year.
Where: click the model name next to the send button, then Effort.
Options: Low, Medium, High, Extra high, Max. One of them is marked Default for the model you chose: use that one for everyday work.
Good to know: on the newest models thinking is always on, so Effort is the only dial. Higher effort reaches your usage limit sooner.
Where: the model picker. On the Free and Go plans there is a Think button instead, which you press for a single message.
Options: Plus has Instant, Medium and High. Pro and Business add Extra High and Pro; Pro runs a stronger model and takes the longest.
Good to know: on paid plans Instant does not switch itself to a higher level, and writing "think harder" in your message does not change the level either. Pick it in the menu.
Where: click the model name at the bottom of the message box. Choose the model first, then Thinking level.
Options: two separate choices. Model: Flash-Lite, Flash or Pro. Thinking level: Standard or Extended. Deep Think is a third level that needs the Pro model and the AI Ultra plan.
Good to know: Deep Think usually takes a few minutes. Stronger models and higher levels use more of your allowance.
Claude Code: type /effort for a slider, or /effort high to set a level directly. Levels are low, medium, high, xhigh and max. Adding the word ultrathink to a message asks for deeper reasoning on that one turn without changing the setting.
Codex: the setting is called reasoning effort (model_reasoning_effort in the config file), with levels such as low, medium, high and xhigh; which ones exist depends on the model.
APIs: the same dial is a parameter on each request, so a program can use a low level for routine steps and a high one for the hard step.
Simple questions get slower, not better, and you run out of messages halfway through the day. Some models also overcomplicate an easy task when told to think hard.
More thinking cannot guess what you left out. Say who it is for, give the material, say what you want back. How to ask helps more than any level.
Thinking is reasoning, not looking things up. For anything recent, turn on web search or use a search tool such as Perplexity, then check the sources.
A slow answer can still be wrong. Check numbers, names and quotes exactly as you would at any other level.