Why Claude Usage Runs Out So Fast: 9 Hidden Drains

Diagnose Before You Optimize

There is plenty of advice on stretching Claude usage, including our own 15 ways to never hit the session limit and how to extend your session limit. This post is the step before those: working out why your usage disappears, so you fix the actual cause instead of applying every tip at once.

The core fact is simple. Anthropic's help center says how fast you reach a limit depends on the model, how long your chats are, and how large your files are, and that tools, connectors, and extended thinking all draw on your allowance. Everything below is a specific way one of those factors grows without you noticing.

Drains 1-3: The Conversation Itself

1. Long threads re-read their history. Claude does not remember a conversation between messages the way a person does. Each reply is generated with the prior conversation as input. Message 40 in a thread carries the weight of messages 1 through 39. This is why a one-line follow-up late in a long chat can visibly move the usage meter.

2. Automatic context management. When a conversation gets long enough, Claude manages context automatically to keep going. Anthropic's help center states that longer conversations that trigger automatic context management consume more of your usage limit. If your thread has reached that point, you are paying a premium to keep it alive.

3. Retries and regenerations. Every regenerate is a full new response with the full context behind it. Five retries of an answer near the end of a long thread can cost about as much as five new questions in that thread. Editing your earlier message to steer the answer is usually cheaper than regenerating the same prompt repeatedly and hoping.

Related reading: why long Claude conversations lose context covers the quality side of the same problem.

Drains 4-6: What You Attach

4. Large files stay in the thread. An uploaded PDF or spreadsheet is not read once and forgotten. It becomes part of the conversation context and is carried with later messages. Claude's API documentation estimates 1,500 to 3,000 text tokens per PDF page, plus image processing for each page when visual analysis applies. A 60-page report can easily be the largest thing in your thread.

5. Images and screenshots. Images are processed as visual tokens. Per Claude's vision docs, a 1000x1000 image costs about 1,296 visual tokens, and larger images cost more before being downscaled. Ten full-resolution screenshots in one debugging thread add up.

6. Oversized project knowledge. Project files are available to every chat in the project. A project stuffed with every document you own adds cost and noise to chats that only needed one of them. Remove files you no longer use. Our guide to how files count against Claude usage goes deeper on trimming.

Drains 7-9: Features and Settings

7. Extended thinking left on. Thinking produces reasoning tokens before the answer. That is worth it for hard analysis and wasted on "reformat this as a table." Anthropic's help center suggests turning it off when you do not need enhanced reasoning.

8. Tools, connectors, and web search. Anthropic describes tools and connectors as token-intensive. Search results, connector responses, and tool outputs are pulled into context. The help center recommends turning off connected apps you do not need and asking Claude not to search the web when you do not need current information. Research-style tasks that run many searches are among the most expensive things you can ask for.

9. Model and effort level. The strongest model at high effort is the most expensive way to answer any question. Anthropic recommends lower effort for routine tasks. Heavy Claude Code use also counts here, since it shares the same limits as claude.ai.

A Diagnosis Table

Match your symptom to the likely drain. This is faster than guessing.

Worked Example: One Question, Two Costs

Consider the same question asked in two places: "Which of these three vendors has the shortest contract term?"

Before: asked as message 52 in a thread that already contains three 40-page vendor PDFs, a pasted email chain, ten rounds of drafting, web search on, and extended thinking on. Using the API docs' per-page estimate, the PDFs alone could represent well over 150,000 tokens of text before counting the rest of the thread. The answer is one sentence; the input behind it is enormous.

After: asked in a new chat with a brief and only the three contract-term clauses pasted in:

Context: comparing three vendors for a 12-month analytics contract.
Below are the "Term and Termination" clauses from each contract.
Web search is not needed.

Vendor A: [paste clause]
Vendor B: [paste clause]
Vendor C: [paste clause]

Which vendor has the shortest minimum term, and are there
auto-renewal traps in any of them? Answer in a 3-row table.

Same answer, a small fraction of the context. The tradeoff is real: you need to know which clauses matter. If you do not yet know where the answer lives, keep the full documents, ask once, and then move to a lean thread for follow-ups.

A 5-Minute Self-Audit

Run this the next time a session ends sooner than expected:

1. Open Settings > Usage and note whether the session or weekly limit is the constraint.
2. Count messages in your heaviest thread. Past a few dozen long turns, start a fresh chat with a summary.
3. List attachments in that thread and project. Remove anything you are not actively asking about.
4. Check whether extended thinking, web search, and connectors were on for routine tasks.
5. Check whether a Claude Code session ran earlier in the same window or week.
6. Pick the one or two drains that explain most of it. Fix those first.

When not to bother: if you hit the limit once on an unusually big day, the audit is not worth it. It pays off when the pattern repeats.

Also resist the urge to fix everything at once. Turning off thinking, search, and connectors while splitting every thread will make usage last longer, but it can also make answers worse on the tasks that needed those features. The goal is to stop paying for capacity you are not using, not to strip Claude down to its cheapest mode for every job. Keep the expensive features for the tasks that earn them.

Where ClaudeKit Fits

Diagnosis needs numbers. ClaudeKit's live token counter shows a word count and token estimate while you type and estimates on Claude's responses (roughly within 10%), so you can see when a pasted block is the expensive part. The usage circle in the bottom-right of claude.ai shows session and weekly limits on hover when Claude returns usage data.

For drain 1, the Handoff Brief turns a long thread into a compact brief to continue in a new chat, and pinning key messages into the Workspace panel keeps decisions findable without scrolling the old thread. See moving a Claude conversation to a new chat for the full workflow.