- Claude, ChatGPT and Gemini all limit paid plans in a five-hour window and a weekly allowance. Hitting the five-hour limit early usually means a few heavy habits, not too much work.
- Long chats are the biggest drain: every new message makes the AI re-read the whole conversation. In our example, a 40-turn chat reads 3.5 times as many tokens as the same work split into four fresh chats.
- The model, the thinking or effort level, connected tools and large files all use more of the allowance. Anthropic, OpenAI and Google say so in their own help pages.
- The fixes are habits: start a fresh chat per task with a short handover note, use the lighter model for routine work, lower the effort level, and switch off tools you are not using.
You pay for Claude Pro, ChatGPT Plus or Google AI Pro, and by mid-afternoon you are told to wait a few hours. It is tempting to think you simply use AI a lot. Usually something else is going on: a few habits make each message far more expensive than it needs to be. This guide explains how the three main assistants count usage today and the changes that make the same plan go further.
How the limits work now
- Claude: paid plans have a session limit that resets five hours after your first message, and weekly limits across all models, with a separate weekly limit for Fable on some plans. Usage in claude.ai, Claude Code and Claude Desktop comes out of the same allowance. You can see what is left under Settings, then Usage.
- ChatGPT: for Work and Codex, Plus and some Business and Pro plans have a five-hour window and a weekly window, and you need allowance left in both. OpenAI's estimates for Plus range from about 5 to 45 local messages per five hours on GPT-6 Astra to 350 to 3,000 on GPT-6 Luna, so the model you pick changes your allowance many times over.
- Gemini: Google's limits are based on compute, not a fixed number of prompts. They refresh every five hours until you reach a weekly limit. AI Plus gets twice the standard limits, AI Pro four times, and AI Ultra five or 20 times AI Pro. Settings, then Usage Limits, shows where you stand.
All three say limits can change without notice, so treat any number as a guide.
Why long chats eat your limit
An AI assistant does not remember a conversation the way you do. Each time you send a message, it reads the whole conversation again from the start, then writes its reply. The fortieth message in a chat makes it re-read the thirty-nine before it. Google says it directly: long chats cost more because Gemini must re-read the conversation history. Anthropic says longer conversations use more of your limit, and those that trigger automatic context management use even more.
Here is how fast it adds up when each turn, your message plus the reply, adds about 1,000 tokens:
| Same amount of work | In one long chat | Split into fresh chats of 10 turns | Long chat reads |
|---|---|---|---|
| 20 turns | 210,000 tokens read | 115,000 (2 chats) | 1.8× as much |
| 40 turns | 820,000 tokens read | 235,000 (4 chats) | 3.5× as much |
| 60 turns | 1,830,000 tokens read | 355,000 (6 chats) | 5.2× as much |
The split chats each start with a 500-token handover note so nothing is lost. Providers do not publish exactly how re-reading is weighed against your allowance, so read these as the shape of the problem rather than an exact count: the cost of a chat grows with the square of its length, and splitting it grows only in a straight line.
What else uses more
- The model. The strongest model uses the most for the same task. Most messages, such as approvals, short questions and quick edits, do not need it.
- Thinking and effort. Anthropic and OpenAI both say higher effort uses more, and OpenAI adds that it does not always improve results.
- Tools and connectors. Anthropic calls them token-intensive: each connected app adds to every message, even when you do not use it. Web search pulls whole pages into the chat.
- Large files. Long PDFs, spreadsheets and codebases are read again with the rest of the chat. Google rates them as medium cost; Deep Research and video generation as high.
- Fast modes. OpenAI says Fast and Ultrafast draw your allowance down faster.
Eight ways to make your plan last
- One task, one chat. When the topic changes, start a new chat. Anthropic and Google both recommend it in their help pages.
- Carry a handover note, not the whole history. Before you leave a long chat, ask for a five-line summary of the goal, the decisions made and what is left, then paste it into the new chat.
- Use the lighter model for routine work. Save the strongest model for the hard part of a task. In Claude Code, a lighter main model with a stronger advisor does the same at key decisions only.
- Lower the effort level, or turn thinking off, for simple questions, and raise it only when an answer is not good enough.
- Approve in one go. "Yes, and also do X and Y" in one message costs one re-read; three separate replies cost three.
- Switch off tools and connectors you are not using, and tell the assistant not to search the web when it does not need current information.
- Keep project instructions and files lean. Remove files you no longer need; they are read with every message in that project.
- Save repeated requests as instructions. If you paste the same long brief every week, put it in a project or custom instruction once instead.
Find out what is using yours
Everyone's usage leaks in different places. Some people lose most of it to two or three very long chats, others to the top model answering one-word approvals. AiGanak Scan reads your Claude, ChatGPT or Gemini export and shows how much of your usage went to re-reading long chats, which messages did not need the top model, and the fixes ranked by how far they would stretch your plan. You can also paste a single chat for an instant result. It is free and private by design: your chats are never stored or shared, and no one at AiGanak sees them.
Building with the API rather than a chat plan? The same ideas, in the form of caching and batch discounts, are in our guide to reducing LLM API costs.
Sources
- Anthropic: How usage and length limits work (checked 20 November 2026)
- OpenAI: Managing usage with GPT-6 Astra in Work and Codex (checked 20 November 2026)
- Google: Gemini Apps limits and upgrades for Google AI subscribers (checked 20 November 2026)
Provider prices and limits change often. Figures were checked on 20 November 2026; confirm current terms with each provider before you rely on them.