How Claude tokens
actually work.
Four different things get called "the limit" in the same conversation, and they behave differently. Sorting them out takes about four minutes and saves a lot of arguing with a progress bar.
A token is a chunk of text, usually a short word or part of a longer one, and it is the unit everything gets measured in: what you send, what comes back, and what sits in the middle. The confusion is not really about tokens, though. It is that two separate limits both get described as running out, and the fix for one does nothing for the other.
Usage limits and length limits are not the same thing
Anthropic draws the line clearly: usage limits control how much you can use Claude across all your conversations, while length limits control how long any single conversation can become. Hitting a usage limit means waiting for a reset or adding credits. Hitting a length limit means this particular thread is full. Starting a new chat fixes the second and does nothing at all for the first.
The context window is a desk, not a memory
Everything Claude can see in one conversation has to fit on that desk at once: your instructions, the files you attached, everything either of you has said so far. On paid plans the newest models support up to a one million token window, while others support 500,000 or 200,000. Nothing is remembered between conversations unless you deliberately put it back on the desk, which is why the second conversation about a document needs the document again.
Cost is not linear, because the desk is re-read every turn
This is the part that surprises people. Each new message is not priced on its own; the whole conversation goes back through on every turn. A one-line question at the end of a very long thread is an expensive one-line question. That is why usage seems fine all morning and then disappears in an hour. The cost of a message is set by everything that came before it, not by its own length.
Attachments, tools and effort all move the dial
Anthropic lists the factors that affect usage plainly: message length, file attachment size, current conversation length, tool usage, model choice, effort level, and artifact creation. Tools and connectors in particular are described as token-intensive. None of that is hidden, and almost nobody reads it before wondering where their allowance went.
Caching is the lever most people never pull
Upload a document to a project and it is cached for future use, so every time you reference it, only the new or uncached portions count against your limits. The practical version: anything you will refer to more than twice belongs in a project rather than pasted into a fresh chat each morning. It is the cheapest single change most people can make.
Start a new conversation more often than feels natural. Long threads are the single biggest driver of consumption that is entirely within your control, and the instinct to keep everything in one place is exactly the wrong one. If a task needs history, put the history in a project so it is cached, rather than in a thread that gets re-read on every turn. Our pricing page sets out how we work with teams on this properly.
Things people
usually ask.
Roughly speaking a token is about three-quarters of an English word, so a thousand tokens is around 750 words. Treat that as a rule of thumb rather than a formula, because different models count the same text differently.
It gives you a fresh context window, not a fresh usage allowance. Those are separate limits. A new conversation is cheaper per message because there is less history to re-read, which helps, but it does not reset the cap you are working against.
Output is priced higher than input on the API, and the same asymmetry shows up in how quickly subscription usage disappears. Asking for a summary rather than a full rewrite is a real saving, not a cosmetic one.