Claude and ChatGPT limits: where your tokens go.
Your limit is not spent in messages but in tokens: on every reply the model reads the whole conversation again. Here is how that works, how much the context you give your AI really weighs, measured on a real project, and what you can do to spend less.
What eats your limit
From Anthropic's official pages, read on 10 October 2026.
A model does not «know» your things: it reads them, every time, inside the context window, its working memory for a single chat. And it does not read them once. Claude Code’s documentation puts it this way: the full conversation goes out with every request, and a one-line question in a session that has been open all day still draws usage for the whole conversation. The parts already read are re-read at the cached token rate, but they are re-read.
Claude’s help center, as of 10 October 2026, says usage depends on the length and complexity of your conversations, the features you use and the model. Paid plans have a session limit that resets every five hours and a weekly limit, and claude.ai, Claude Code and Claude Desktop all draw from the same limit. The window reaches 1 million tokens on the newest models, 500,000 or 200,000 on the others, and part of it is always kept for the reply. Tools and connectors, Anthropic writes, are token-intensive.
How much cached tokens weigh against the limits, Anthropic does not publish, and we will not make it up: below we count tokens, not percentages of your plan. ChatGPT too has a context window and per-plan limits: the figures are on OpenAI’s plan pages. The details are in our articles on Claude usage limits and on ChatGPT vs Claude limits.
How much it weighs, measured
A real project, with Claude's own token count.
The example is the Cobrain project in our own brain: 134 notes. The comparison is between two ways of giving the AI its context at the start of a chat. Without Cobrain: the files the project declares always necessary (card, index, decisions), handed over in full. With Cobrain: the summary the AI receives when it opens the project, plus the connector’s fixed weight.
| First message | Without Cobrain | With Cobrain |
|---|---|---|
| Project context | ≈69,400 | 4,700 |
| Connector tools | — | ≈5,200 |
| Total, in tokens | ≈69,400 | ≈9,900 |
On the first message that is 7 times fewer tokens.
The fixed cost, stated up front
An MCP connector is not free: the definitions of its tools enter the context too. Cobrain’s weigh 5,200 tokens when the app loads them in full (measured with Claude Code’s /context, with Cobrain on and off), and they are already counted in the total above. Claude Code, according to its documentation, loads only the tool names and the server instructions by default, and the rest when a tool is needed.
Over a whole chat
We took 35 real working chats on the same project, with every note the AI opened. Over the whole chat Cobrain uses 2.9 times fewer tokens (median). Counting that the model re-reads the context on every reply, 3.6 times fewer (median).
Where it costs more
In 4 chats out of 35 Cobrain brought in more tokens than «without». They are chats that opened many notes: notes you would have had to hand the AI anyway without Cobrain, and that our «without» does not count, because it stops at the three base files.
The method
We counted tokens with Claude’s own count, derived from 258 real Claude Code conversations (the input between two consecutive turns), cross-checked with OpenAI’s tokenizer. On these notes, Italian plus Markdown, Claude counts about 2 characters per token: on the same texts, about 1.58 times the tokens OpenAI’s tokenizer counts. Anyone can redo it on their own notes: in Claude Code /context shows what fills the context and /usage how much you have used. Try the same question with the files pasted in and with the memory connected.
How Cobrain serves it
In layers, saying what it left out.
Rules first, then the card, then the notes
A folder with an index card is a unit: a client, a project. When your AI opens it, it gets the brain’s rules, the card, the lessons learned and the latest decisions. The other notes it searches for and opens only when needed.
Whole lines, never fragments
Search returns whole lines, not pieces cut in half: one fragment once turned «Alessandro’s sister is called Alessia» into «Alessandro is called Alessia». And the AI is told to read the note before answering.
A stated ceiling
When it loads a large branch, Cobrain stops at a limit and says so at the top: the AI knows it has not read everything, and can ask for the rest.
The brain is never dumped whole
A brain of six hundred notes never enters a chat in full. Connecting it takes two minutes: see the guides for Claude and for Claude Code.
What you can do
Habits that pay, with or without Cobrain. From the official guides, as of 10 October 2026.
- A new chat for every new job. A long chat is re-read on every message. In Claude Code,
/clearbetween unrelated tasks: stale context, the guide says, wastes tokens on every message that follows. - A short CLAUDE.md. It loads at the start of every session. The guide suggests staying under 200 lines and moving instructions for specific workflows into skills, which load only when invoked.
- Turn off connectors you are not using, and web search when you do not need it. In Claude Code,
/mcplists the connected servers. - Lean projects. Short instructions, and remove the files you no longer need.
- Less reasoning for simple things. Lower the effort, or turn off extended thinking, for routine tasks.
- Specific requests. «Improve this codebase» makes it read everything; «add input validation to the login function in auth.ts» makes it read one file.
If you work in Claude Code, read our article on how to reduce Claude Code token usage. The tools the AI uses on Cobrain are documented in full.
Quick answers
Why does Claude hit its limit so fast?
Because on every reply it re-reads the whole conversation, files and tool results included. The longer the chat, the more each message costs, even a short one. New chats, connectors turned off when unused and a short starting context make the difference.
Does a memory MCP server use tokens?
Yes, in two ways: its tool definitions and the notes the AI reads. Cobrain’s definitions weigh 5,200 tokens. It pays off when it saves you pasting heavier files: in our measurement that happened in 31 chats out of 35, not all of them.
How big is Claude's context window?
According to Claude's help center, as of 10 October 2026: up to 1 million tokens on the newest models, 500,000 or 200,000 on the others, with part of it kept for the reply.
Does Claude's limit reset every 5 hours?
Paid plans have a session limit that resets every five hours, plus a weekly limit. claude.ai, Claude Code and Claude Desktop share the same limit. More in our article on Claude usage limits.
Does this apply to ChatGPT too?
ChatGPT also has a context window and per-plan limits, but the figures are OpenAI’s to publish and we do not repeat them here; the comparison is in our article on ChatGPT vs Claude limits.
Are my notes safe if I hand them to the AI this way?
On Cobrain notes and attachments are encrypted, with one key per brain: the database cannot read them. Cobrain does not change what Claude or ChatGPT do with what they read, though: that depends on their terms and settings. Details in permissions and privacy.
Give your AI only the context it needs.
An index card, the decisions, the rules: your AI reads them at the start of a chat and opens the rest only when needed. You can start for free.
