ChatGPT Slider and Think Button: How to Use Both

There is a control next to your ChatGPT message box that was not there last week, nothing on screen explains what moving it does, and you have no idea whether pushing it up costs you anything. Here is the shortest route through it.

OpenAI added the reasoning-effort slider in its August 6, 2026 ChatGPT update, along with a Think button, an updated GPT-5.6 Sol behind paid conversations, and GPT-5.6 Luna as the model answering free accounts. The same OpenAI post says Free and Go accounts get unlimited text chats.

Both controls do one job at different granularities. The slider sets how much reasoning the model does by default, message after message. The button forces the deeper path for a single message and leaves your default alone.

The path on Plus and Pro

Paid consumer accounts get both controls, per OpenAI's August 6 announcement. The sequence that gets you a useful result on day one:

1. Update the ChatGPT app and reopen it. On the web, hard-reload the tab. If nothing new appears after that, the rollout may not have reached your account yet.
2. Open a new conversation and look at the composer, not Settings. These are in-chat controls, so they belong with the message box rather than behind a preferences menu. Hunting through your account settings for a switch is the most common way to waste ten minutes on this.
3. Check which model is answering. After the update, paid conversations run on the updated Sol and free accounts get Luna. If a long-running chat still shows an older selection, opening a fresh conversation is faster than trying to convert the old one.
4. Set the slider before you send, not after. It governs the message you are about to submit. Changing it once an answer is on screen does nothing to that answer, and you have to send again to see the difference.
5. Leave the slider mid-range as your standing default. Then reach for the Think button on the individual messages that deserve it. A permanent high setting turns every "what's a good subject line" into a slow one.
6. Read the pause as feedback. Higher effort means the model generates intermediate steps before the first visible token, so the wait before text appears grows. A long pause followed by a structured answer means the setting took. An instant reply to a genuinely hard question means it did not.

Neither control changes which model you are talking to. They adjust how much work the model you already have does on one message. Switching models is a separate decision made in the picker, and confusing the two is how people end up convinced the slider is broken.

The free-account path

Free and Go accounts get the Think button and no slider, per the same OpenAI announcement. That makes the procedure shorter and, in one respect, better.

1. Write the message first, at whatever length it needs. Effort applied to a vague question produces an elaborate answer to the wrong question.
2. Tap Think, then send. It applies to that message.
3. Use it on the messages where a later step depends on an earlier one being right. Multi-step arithmetic, a plan with ordering constraints, a comparison with more than three variables, code where the second function consumes the first one's output.
4. Skip it on lookups, rewrites and tone changes. Reasoning depth buys nothing when the task has no chain in it.

The unlimited-text-chat change matters more here than the button does. When messages were rationed, spending one on an experiment felt expensive. It no longer does, so the cheapest way to learn what the button is worth is to send the same hard question twice, once with it and once without, and compare the two answers step by step rather than by how confident each one sounds.

Effort settings by task type

The slider ships without named modes to work from, so the table below reads it as low end, middle and high end. This is a recommendation rather than a documented mapping, and the thing it sorts on is simple: whether the task contains steps that depend on each other.

Failure modes

Neither control appears. Looks like: a composer identical to last month's. Update, relaunch, start a new chat. If a colleague on the same plan has it and you do not, the rollout has not reached your account, and nothing in your settings will speed that up.

The slider is pinned high and everything feels slow. Looks like: multi-second pauses before trivial replies. Reasoning tokens are real tokens even when the interface hides most of them, so depth costs you time on every message you send, including the ones that were never going to benefit. Drop back to the middle and let the button carry the exceptions.

The slider is pinned low on work that needed the depth. Looks like: a fast, confident, well-formatted answer that skips a step. This one is dangerous because the output quality signal is inverted. A shallow answer to a hard question reads better than a careful one, right up until you check the third bullet.

You changed the setting after sending. Looks like: an unchanged answer and the suspicion that the control does nothing. Send the message again with the new setting rather than assuming it applies retroactively.

You used Think to fix missing information. Looks like: a longer, more elaborate answer with the same wrong fact in it. Effort improves how the model works through what it has. It does not import anything it lacks. When the failure is a missing fact, attach the document, turn on search, or paste the source, and leave the effort setting where it was.

You are on Free and looking for the slider. It is not there. The August 6 update gives Free and Go the button, and our tier-by-tier breakdown has the full split. Per-conversation depth as a standing default is what the paid tiers sell; the button is a per-message decision you make each time.

You expect the update itself to feel different. Looks like: a week of ordinary chats that feel exactly like last month's. What landed underneath the controls on August 6, per OpenAI's post, is a change to which model answers which tier and the removal of text-chat limits for Free and Go. You notice the second one at the end of a long session you would previously have been cut off in, not in any single reply.

Wrong tool territory

Reasoning depth is a lever on thinking, and a large share of what people ask a chatbot to do is not thinking-limited at all.

When the answer depends on something that happened this week, the constraint is retrieval. Search, a pasted source, or a tool built for the job like Perplexity beats any slider position. A model reasoning harder over facts it does not have will hand you a more confident version of the same error.

When the job is a long agentic run across a repository, the chat window is the wrong container regardless of effort. That work belongs in Codex, Claude Code, or an IDE agent that can read files, run tests and iterate without you shuttling text back and forth. Turning the slider up does not give the chat window file access.

Press Think when you can name the step you expect the model to get wrong. If you cannot name one, and your ChatGPT use is drafting, summarising, translation and questions you already know enough to check, leave the slider where it shipped, ignore the button, and put the attention into the prompt instead.