Last time I wrote about pasting far too many images into one chat with Claude and hitting that conversation’s limit. My fix was simple: start a fresh chat. But once I was in the new one, I had a nagging feeling.
I moved to a new chat, so why does my usage still feel this fast?
My first thought was that I was imagining it. But it kept bothering me, so I asked my AI partner Kuro (Claude) straight out: is this just in my head?
The short answer: no. There is a real mechanism behind it. So this time I want to write down what a “token” is, the way I needed it explained. Once it clicks, working with AI gets a lot more comfortable.
A token is the counter AI uses when it reads and writes
The most important word first. A token is the basic unit AI uses when it processes text — roughly, a character counter.
In Japanese, one character comes out to around one to two tokens. And AI spends tokens on both sides: the text I send in, and the text it sends back.
There is also a ceiling on a single conversation: in the web version of Claude, one conversation can handle up to about 200,000 tokens.
Picture that as a big bucket. The longer the conversation runs, the more of what you and the AI have said piles up in it.
The key part: images eat a lot of tokens
Here is the real point. Tokens are a counter for text — but images consume tokens too. Quite a lot of them.
Claude reads an image by cutting it into small squares: picture it divided into little blocks of 28×28 pixels, with the AI looking at each one. Every one of those blocks counts as a token.
A large image, or a detailed screenshot, contains a great many squares — so one of them alone can consume a great many tokens.
As I wrote last time, I had pasted 83 screenshots into one chat. So it was not only my text in that bucket — the tokens for all 83 images were in there too, weighing it down. When I went looking, it was stated plainly:
Just uploading files consumes tokens, so if you handle a large number of them, you will reach the limit sooner.
My gut feeling had been right.
Why the same chat keeps getting heavier
One more mechanism is worth knowing. Claude remembers the images and files you uploaded for as long as you stay in that same chat. That is genuinely useful: if I want to ask about an image I pasted earlier, I do not have to paste it again.
But turn it around: as long as I stay in that chat, I am carrying those 83 images on my back through every exchange.
Each time I send a message, the AI re-reads the whole conversation so far — the 83 images included — before working out its reply. That is why the longer a conversation runs, the heavier each round trip becomes.
So a new chat is lighter — but only up to a point
Once you see that, my feeling starts to make sense. Moving to a new chat resets the bucket to empty, and for a while things are light again. That part is true — you put down the weight of those 83 images.
But then you start working there. You paste in a handover note, then fresh screenshots. Those tokens get spent all over again.
On top of that, Claude’s usage limits are not decided by a simple count of messages sent. How much you consume depends on the length and complexity of each message and the size of the files you attach. Work that leans on images is heavier per message from the start.
So “I switched chats and my usage still goes fast when I work with images” was not imagination. It was the mechanism doing exactly what it does.
Three habits that make it easier to live with
Here are the three things I actually do.
Start a new chat at a natural break in the work
Rather than dragging along a chat that has grown long and heavy, I move to a new one at a clean stopping point. It runs lighter and burns less, and with a handover note ready, picking the work back up is smooth.
Cut the images down to the one you need
Pasting fewer images out of habit makes a real difference to token consumption. Instead of several shots of near-identical screens, pick the one that shows the point. That alone pushes the limit much further away.
Send your questions together
This is less about saving tokens than saving messages — but if you have several things to ask, sending them in one message costs less than dribbling them out one at a time.
What I take away from it
- A token is the counter AI uses when it reads and writes.
- Images are recognised by being cut into squares, so pasting them eats a lot of tokens.
- Within one chat you keep carrying every past image, so the longer the conversation, the heavier it gets.
- A new chat resets that and feels lighter — but paste images there and of course you spend again.
- So “it feels like it got heavier” is not imagination. It is the mechanism doing what it does.
For someone not good with computers, “token” sounded like a difficult word at first. But once you know it is a counter for the characters AI reads and writes, and that images eat a lot of them, there is nothing to it. Understanding how something works makes the AI a much more relaxed partner.
One note on the plan I use
Part of the reason I can keep long sessions going with Kuro is that I pay for the Pro plan. The mechanism is the same on the free version, but work that uses a lot of tokens and images does hit the limit sooner there. If you want to sit down and work through something properly, I would go with the paid plan.
Next time I plan to show how I write the “handover note” I mentioned here, with a real example.
For background, I wrote about the day I filled a chat with screenshots in what to do when Claude says the image limit is exceeded, and about which work burns usage fastest in three tasks that eat AI usage.
