You send a short follow-up message to Claude. Four words. Something like "make it more concise." And somehow that tiny message eats 28,000 tokens — nearly as much as your entire usage limit for the day.
That's not a bug. It's how Claude works. And once you understand the mechanism behind it, you'll completely change how you use the tool.
The Real Token Killer: Conversation History
Most people assume tokens are consumed by what they send. Write a long prompt, spend a lot of tokens. Send a short message, spend very few.
That's wrong.
Every single time you send a message in an active Claude conversation, Claude doesn't just read your new message — it re-reads the entire conversation from the beginning. Every message you've sent. Every response it has given. All of it, every time.
This is how large language models maintain context. They don't have memory in the traditional sense. They maintain coherence by processing the full conversation history with every new exchange. The longer your thread gets, the more tokens each new message costs — not because of what you typed, but because of everything that came before it.
A first message in a fresh thread: a few hundred to a few thousand tokens.
A follow-up message on a thread with 20 exchanges: potentially 20,000–40,000 tokens, even if your actual message is five words.
A Real Example: The Exponential Scaling Problem
Here's what this looks like in practice. Someone asks Claude to help create a PowerPoint presentation. The initial request costs around 6,000 tokens — reasonable for a detailed task.
They send a follow-up asking for one small adjustment. Same thread, short message.
That follow-up costs nearly 30,000 tokens.
The follow-up wasn't five times more complex. The follow-up was simpler. But the thread was now five times longer, and Claude re-processed all of it.
This is the core trap that catches most Claude users off guard. They don't hit their limit because they asked for something big. They hit their limit because they stayed in the same thread too long.
The Projects Feature: Convenient But Costly
Claude's Projects feature lets you attach files and documents that persist across conversations — a great idea in theory. Upload your research notes, a business brief, or reference material, and Claude has access to them across sessions without you pasting them in every time.
The hidden cost: every attached file becomes persistent context. Every conversation in that project starts with Claude reading all of it, every time.
Attach a 100-page document to a project and every single message in that project — even a one-line question — costs tokens proportional to that document plus your conversation history.
This catches heavy users particularly hard. A project that seemed like an organisational convenience turns into a token furnace because every interaction is carrying the weight of an entire document library.
How to Track What You're Actually Using
Claude's free plan doesn't show token usage by default. To see it, you need to go to Account Settings → Usage — a menu most users never visit.
A more practical solution is the Claude Counter browser extension, which adds a real-time token counter to the Claude interface. It shows you exactly how many tokens each message costs, so the invisible becomes visible.
Once you can see usage in real time, the conversation history problem becomes immediately obvious. Watch the counter jump between a fresh thread and an extended one and the exponential scaling is impossible to miss.
5 Things That Drain Your Tokens Without You Realising
1. Long Threads on Complex Tasks
Every exchange adds weight to every future exchange. A thread you've been working in for an hour is far more expensive per message than a fresh one — even if the task is simpler.
2. Project Files You No Longer Need
Old documents attached to a project continue consuming context every session. A research file from three weeks ago that you never removed is still adding tokens to every message you send in that project.
3. Uploading Large Documents Unnecessarily
Pasting a 50-page PDF to ask one specific question burns tokens proportional to the full document. Extract and paste only the relevant section instead.
4. Back-and-Forth Refinement in One Thread
"Make it shorter." "Now add a section on X." "Change the tone." Each of these short messages carries the full weight of the growing thread. Five rounds of refinement in one thread can cost more than starting fresh each time.
5. Staying in a Thread Out of Habit
Many users keep adding to existing threads simply because it feels more organised. But there's rarely a technical reason to stay in the same thread unless you genuinely need the full prior context for the next task.
How to Use Claude Without Burning Your Limit
Start New Conversations Deliberately
Before sending a message, ask: does this genuinely need the context from the current thread? If the answer is no — or even "sort of" — start a new conversation. You'll do the same task for a fraction of the tokens.
This is the single highest-impact habit change for heavy Claude users.
Summarise Before You Continue
If you do need to carry context forward, don't just keep going on the same thread indefinitely. At a natural stopping point, ask Claude to summarise the key decisions, outputs, or context from the current session. Then paste that summary into a new thread instead of carrying the full history.
A 200-word summary of a 10,000-token conversation costs a fraction of what continuing the thread would cost.
Audit Your Project Files Regularly
Once a month — or whenever you notice unusual token consumption — go through your Claude Projects and remove files you no longer actively need. Anything that was useful for a one-time task and isn't genuinely needed for future work should be deleted from the project.
Be Surgical With Document Uploads
When you need Claude to reference a document, don't upload the whole thing unless the whole thing is relevant. Copy and paste the specific section, paragraph, or data you actually need answered. You'll get a more focused response and spend far fewer tokens.
Use Shorter, Denser Prompts
This matters more in long threads than in fresh ones, but it adds up. Verbose, conversational prompting in a long thread amplifies the cost. Write clearly and concisely.
Does This Apply to Paid Plans Too?
Yes. The token mechanics are identical whether you're on a free or paid plan. Paid plans give you a higher limit, but the same exponential conversation history problem applies. Heavy paid users hit their limits for the same reasons free users do — they just have more runway before it happens.
Understanding token consumption is valuable regardless of your plan. Efficient usage means you get more done within any limit.
The Mental Model Shift
The useful way to think about this: Claude conversations aren't like chat threads where older messages sit quietly in the background. They're more like a document that grows with every exchange — and Claude reads the entire document every time you ask it anything.
Once that clicks, the right behaviour becomes intuitive:
- Shorter threads for focused tasks
- New threads when context doesn't carry over
- Leaner project file libraries
- Summaries instead of endless continuation
None of this means being paranoid about every message. It means being deliberate about when long context actually helps you versus when it's just burning through your limit for no benefit.
Quick Reference: Token-Saving Habits
| Habit | Impact |
|---|---|
| Start a new thread when context isn't needed | Very high |
| Summarise then restart for long projects | High |
| Remove unused files from Projects | High |
| Paste only relevant document sections | Medium |
| Write concise prompts | Medium |
| Check usage regularly via Account Settings | Low (awareness only) |
Claude is one of the most capable AI tools available in 2026. Using it well means understanding how it actually works — not just what it can do. Explore more AI tools and guides at the Humbaa AI tools directory. Related reading: What Are AI Models and Facts About AI.