7 Ways You Are Wasting Claude Tokens (And Getting Worse Answers For It)

Quick definition first. A token is a small chunk of text, roughly three-quarters of a word. Claude counts tokens, not messages, when it bills you or limits you.

The real problem is not just cost. Wasted tokens also make Claude’s answers worse. You end up paying more for a weaker result.

Table of Contents

Start With The Easy Wins

You have probably seen these before. They work, they take seconds, so do them and move on.

• Start a new chat when you change topic. Claude re-reads the whole conversation every time you send a message
• Say what format you want upfront. “One paragraph” or “just the code” saves a lot
• Pick a smaller model for simple jobs. You do not need the most powerful one to fix a typo
• Turn off extended thinking, web search, and connectors when you are not using them

Now for the seven things that actually cost you the most. The first four apply to everyone. The last three matter most if you build on the API.

1. You Are Uploading Files The Expensive Way

Most people attach the same reference document to chat after chat. You pay for it in full, every single time.

Anthropic’s own help pages say files stored in a Project are cached and do not count against your limits when reused.

One user’s experience shows the gap clearly:

• Their 175,000-token master prompt broke a chat on the very first message
•  Placed in Project files instead, it used about 10 percent of the project space and worked fine
•  Same file, same model, completely different cost based only on where it lived
• File type matters just as much:
• One PDF page costs 1,500 to 3,000 tokens, because Claude reads each page as an image plus its text
•  A 20-page PDF can use 30,000 to 60,000 tokens in a single upload
•  A full screenshot costs around 1,500 tokens. A tight crop can drop under 100
• Plain text and markdown cost the least, with no conversion at all

What to do: copy text out of PDFs before uploading, crop screenshots tight, and put anything you will use twice into a Project.

2. Claude’s Replies Cost Five Times More Than Your Questions

Almost every tip out there tells you to shorten what you send. But what Claude sends back costs about five times more per token.

There is also a hidden part. Claude sometimes reasons privately before replying. That thinking is billed at the same high rate, and you never see it.

• Opus 5 costs $5 per million tokens in, and $25 per million out
• Private reasoning often runs 500 to 5,000 tokens before the visible answer starts
• On Opus 5 this reasoning is on by default. On Opus 4.8 it was off unless you asked for it
• The effort setting moved from medium to high, which raises cost without you changing anything

What to do:

•  Ask for the answer, not the explanation, when you do not need the reasoning
•  Set the effort level on purpose. Low is fine for sorting, tagging, and formatting work
• Cap the thinking budget on anything running at volume
• Remember your max_tokens limit covers thinking and the reply together

3. Giving Claude Too Much Context Makes Its Answers Worse

This is the part hardly anyone writes about, and it matters more than the money.

Researchers at Chroma tested 18 leading AI models. Every single one got worse as the amount of text grew. Not some of them. All of them.

They kept the questions equally hard and only changed the length. So the drop came from size alone.

• A 113,000 token conversation history scored about 30 percent worse than a short 300 token prompt on the same questions
•  Adding just one piece of related but unhelpful text pulled results below normal
• Models did better on shuffled text than on well-organized text
• Claude tends to say it does not know, rather than making something up

That last point is useful. If Claude says it cannot find something you know is there, your conversation is too full, not too empty.

Adding more context makes it worse. Anthropic calls this a limited attention budget. Every extra token competes with the ones that matter.

4. You Have Never Checked Where Your Tokens Actually Go

Most people guess. No starting number, no way to tell if a change helped.

Anthropic publishes a useful benchmark. Average cost runs about $13 per developer per active day, and $150 to $250 per month. For 90 percent of users, it stays under $30 a day.

•  /context shows what is filling your space right now, broken down by type
•  /usage and /status show how much you have left and when it resets
• ccusage reads your local logs and shows real cost per session, which matters on flat rate plans
• In the API, cache_read_input_tokens tells you if your caching is working

One developer tracked his usage for a week. Sessions over 100,000 tokens used three times more. Simple questions cost pennies. Leaving windows open cost nothing.

5. Every Connected Tool Costs You Tokens, Even When You Never Use It

MCP servers are connectors. They link Claude to outside apps like GitHub, Notion, or Figma. Useful, but each one has a hidden price.

When you connect a server, Claude loads the full instruction manual for every tool inside it. This happens the moment your session opens, before you type anything.

You pay for that manual whether you use those tools or not.

• The GitHub connector alone uses about 55,000 tokens for its 93 tools
• One developer measured 66,000 tokens already spent at startup, roughly a third of his available space
• A single tool description runs 300 to 600 tokens
• Cursor limits users to 40 tools because more starts causing problems

It gets worse in longer jobs. Claude re-reads that whole manual on every turn. A task with 60 turns pays the same cost 60 times.

What to do:

• Turn off connectors you are not using today. This takes under a minute and almost nobody does it
• In Claude Code, switch on Tool Search. It loads tool details only when needed, and cut one setup from 51,000 tokens to 8,500
•  For bigger builds, Anthropic’s own method loads tools through code instead of upfront. One workflow dropped from 150,000 tokens to 2,000

6. A Date In Your Prompt Can Make Every Request 10x More Expensive

Prompt caching lets Claude save a repeated part of your prompt so you pay far less to send it again. It only works if that part stays exactly the same, character for character.

Change one character and the saving disappears. Nothing warns you. No error appears. You just quietly pay full price forever.

A real production case shows the damage:

• Every request saved 170,000 tokens to cache, but never read a single one back
• Each message cost $0.50 instead of $0.05
• One day of work cost $35 instead of $9
• The culprit was a “Current Date and Time” line in the system prompt

Things that quietly break caching:

• Dates or clock times added to your system prompt
• Session IDs, message numbers, or run details
•  A user’s name or settings placed near the top
•  Version numbers that change every time you deploy
• Tools switched on and off between requests

What to do: put your unchanging content at the top, your cache marker after it, and everything that changes at the bottom.

7. Turning On Caching Can Cost More Than Leaving It Off

Most articles treat caching as free money. It is not. You pay extra upfront before you save anything.

Saving to cache costs 1.25 times the normal rate. Reading it back costs one tenth. So caching only pays off if you actually read it back.

• One save plus one read works out to 0.675 of normal price. A win
• One save plus two reads drops to 0.483
• A save that never gets read costs more than not caching at all
• Below a 50 percent hit rate, caching is losing you money

How to check: send the same request twice within five minutes. If cache_read_input_tokens shows zero both times, your cache has never worked once.

One Old Warning You Can Now Ignore

Many guides still warn that requests above 200,000 tokens cost double. That price jump existed during testing, and Anthropic removed it in March 2026.

Today’s larger models charge the same rate across the full window.

The Pattern Behind All Seven

Waste is rarely in the words you type. It sits in what loads before you start, what never gets read back, and what you never measured.

Run these checks once. Most people find a single fix worth more than every “be more concise” tip combined.