Skip to content

GUIDE · Claude ·

Hit your Claude limit? 4 ways to use fewer tokens, in ChatGPT too

The “limit reached” message usually shows up at the worst moment. Your Claude limit is a usage allowance that resets every 5 hours, and on paid plans every week too. A long chat eats it faster and faster, because with every question, the model receives the whole conversation so far. Below are four ways to turn that around. For each one we say where it works: in the Claude app, in Claude Code, or in ChatGPT.

Time
about 10 min
Steps
4
Cost
$0

In this guide you will learn:

  • why a long conversation uses up your limit faster than a short one,
  • when to open a new chat and when /compact is enough,
  • how to switch to a lighter model and how much that changes the price,
  • how to turn on caveman mode: with a ready-made instruction or the Caveman add-on from GitHub,
  • where to check how much of your limit is left and when it resets.

What you need

  • any plan

    Claude or ChatGPT

    Steps 1, 3 and 4 work in both. Opus is only on paid Claude plans (Pro, Max, Ultra), so switching from Opus to Sonnet applies to paid accounts.

    Claude plans ↗
  • on Pro and above

    Claude Code (optional)

    The /clear command only works in Claude Code in the terminal. If you do not use it, the new chat button is enough for step 1.

    Claude Code commands ↗
  • $0

    Caveman (optional)

    A free add-on from GitHub (Apache-2.0 license, over 118k stars) that makes caveman mode permanent. Without it, our ready-made instruction is enough.

    Caveman on GitHub ↗

Step by step

  1. Step 1: New topic, new chat

    • Why it works: according to Anthropic’s documentation, each message and reply stays in the so-called context window, and previous turns are preserved completely. So at the twentieth question, the model also receives the nineteen before it. Anthropic advises it directly: if you are approaching your limit in a long chat, start a new one.
    • When: you have closed one topic (for example, finished an offer) and are starting another (an email to a client). The old conversation will not help, and every question pays for its length.
    • Claude (browser, app, phone): click “New chat” or “+ New”. In the desktop app the Ctrl+N shortcut works too (Cmd+N on a Mac).
    • Claude Code in the terminal: type /clear. It wipes the conversation and starts fresh. This command is not available in the regular Claude chat or in the desktop app’s Code tab.
    • ChatGPT: click “New chat” in the top left corner.
    • Several questions at once: Anthropic advises grouping related questions into one message. Every separate message means another read of the whole conversation.
    • The same files every week: upload them once to a project (Projects). According to Anthropic, project files are cached, so working with them uses your limit more slowly than uploading them every time. A new chat in the same project still sees them.

    Tip:Before you clear a conversation, save what you need from it (text, decisions, files). A new chat does not remember the old one.

  2. Step 2: /compact: a cheat sheet instead of the whole conversation

    • When: you are in the middle of a long task and the model has to remember what you agreed, so a new chat is not an option.
    • Claude: type /compact. The model summarizes the conversation into a short cheat sheet and keeps working from it instead of the full history. You can add what to keep, for example /compact keep what we agreed about the offer.
    • A good moment: the conversation is going in circles because you are fixing the same thing for the fifth time. A cheat sheet with the decisions works better than dragging along every failed attempt.
    • Automatic in Claude: with code execution turned on, Claude summarizes older messages by itself when the conversation nears its length limit. Anthropic notes, however, that such long conversations use more of your limit, so a manual /compact or a new chat still pays off.
    • ChatGPT: there is no such command. Do it by hand: ask for a summary (ready-made sentence below, in the commands section) and paste it into a new chat.
  3. Step 3: Don’t drive a truck to buy bread: a lighter model

    • Why: according to Anthropic, how fast you reach a limit depends on the model, how long your chats are and how large your files are. The strongest model uses it up fastest, and you do not need it for a short email.
    • Claude (app, browser): click the model name under the input box and choose Sonnet instead of Opus. For very simple tasks, Haiku is enough.
    • Claude Code: type /model and pick a model from the list. In the desktop app the picker sits in the row under the input box, shortcut Ctrl+Shift+I.
    • How much it changes: on the API price list, Sonnet 5.5 costs half of Opus 5.5: $2 instead of $4 per million input tokens and $10 instead of $20 for output. Haiku 5.5 starts at $0.10 for input.
    • A bonus from Anthropic: for routine tasks choose a lower effort level, turn thinking off when you do not need it, and turn off tools and connectors you are not using. Tools use a lot of tokens.
    • ChatGPT: pick a lighter model from the model list at the top of the window. Go back to a stronger one for hard tasks.

    Tip:Rule of thumb: the truck (Opus) for hard problems, the scooter (Sonnet, Haiku) for emails, lists and tidying up data.

  4. Step 4: Caveman mode: fewer words in replies

    • Why: your limit also counts every word the model writes back. Greetings, praise for your question and summaries are pure cost.
    • No-install version (Claude and ChatGPT): copy our instruction from the “Instruction: caveman mode” section below. In Claude, paste it into Settings, in the personal preferences field — it works in every chat. In ChatGPT, paste it into Personalization → Custom instructions. You can also paste it at the start of a single conversation.
    • Version with the Caveman add-on (Claude): it is a free add-on from GitHub by Julius Brussee. In the Claude app or on claude.ai open Customize → Plugins, choose Add → Add marketplace, enter JuliusBrussee/caveman and add Caveman. Then turn on Settings → Capabilities → Code execution and file creation. In a chat, start it by typing /caveman and stop it with “stop caveman”. The author’s instructions: INSTALL.md.
    • Caveman in Claude Code: run claude plugin marketplace add JuliusBrussee/caveman in the terminal, then claude plugin install caveman@caveman. The full set (including compressing logs and files) installs with npx @caveman-ai/cli@latest and needs Node.js 22.13 or newer.
    • What it gives you, honestly: according to the author’s measurements, a plain “answer concisely” instruction does most of the work: Caveman then shortens replies by another 3%, and the stricter /ultracave mode by another 35%. JetBrains measured 8.5% shorter replies and no measurable drop in quality across 86 coding tasks.

    Tip:Caveman mode only shortens replies. It does not reduce what the model reads, so it works best together with steps 1–3.

Instruction: caveman mode

Paste it into Claude’s personal preferences or ChatGPT’s Custom instructions. No installation needed.

> Claude
Answer briefly and to the point, like a matter-of-fact expert.
- Answer first, then the reason in one sentence, then the next step.
- No greetings, no praising the question, no assurances that you are happy to help, no summary at the end.
- Short words instead of long phrases. One idea per sentence, 20 words at most.
- Never drop the words "not", "never", "only". Copy numbers, names, code and commands exactly.
- For security, actions that cannot be undone, or when you can see I am confused, write in full sentences.
- Answer in my language.

Ready-made sentences to copy

Paste into the chat when you need it.

  • Summarize our conversation in 10 points: goal, agreements, decisions and open issues. No intro. I will paste it into a new chat.
  • /compact keep the agreements, decisions and the list of fixes to do
  • Answer in 3 sentences at most.
  • Give me only the finished email text, no comment.
  • No intro and no summary. Facts only.
  • stop caveman

The catches, honestly

  • /clear only works in Claude Code in the terminal. In the regular Claude chat and in ChatGPT, use the new chat button.
  • ChatGPT has no /compact command and no Caveman add-on. What is left is a manual summary and our instruction in settings.
  • We have not installed Caveman ourselves. We describe the author’s instructions, and the author notes that the Claude app route has not been checked end to end yet.
  • The Caveman version installed with npx sends anonymous usage stats by default (never your conversations). Turn them off with caveman telemetry off.
  • Caveman’s savings figures come from the author and the companies cited. Your results may differ.
  • Caveman can also add tokens, because the model reads its rules with every reply. The author advises comparing usage with and without Caveman, and turning it off if it comes out higher.
  • Claude limits reset 5 hours after the first message of a session (session limit), and on paid plans every week too (weekly limit). These methods do not raise your limit; they make you use it more slowly.

Questions and answers (FAQ)

How do I check my Claude limit?

Open Settings → Usage. It shows your usage in the current session, how much time is left and when the limit resets.

How often does the Claude limit reset?

According to Anthropic, the session limit resets in full 5 hours after the first message of that session. Paid plans also have a weekly limit: it resets once a week, at a fixed time set for your account.

Claude stopped answering. Is it the limit?

If you see a usage limit message, yes: your session or weekly limit has run out and you have to wait for the reset. A conversation length message only applies to that one chat, so a new chat or /compact will help.

I typed /clear in the Claude chat and nothing happened

That is expected. /clear is a Claude Code command for the terminal. In the chat, click “New chat” or press Ctrl+N in the desktop app.

Is Sonnet worse than Opus?

For hard, multi-step tasks Opus is stronger. For emails, lists, summaries and tidying up data, Sonnet is more than enough and uses less of your limit.

I don’t see Opus in the model list

Opus is only on paid Claude plans (Pro, Max, Ultra). On the free plan you are already working with a lighter model.

After a while the model writes long replies again

Remind it of the rules in one sentence, or paste the instruction into settings rather than the chat. Settings apply to every new chat.

Want your whole team to work with AI for less?

We will show you how to set up Claude and ChatGPT in your company so limits last longer and answers stay to the point.

Sources

As of 2026.11.07.

Related articles

← All Academy resources