How to Cut Your Claude Bill

If your Claude usage is adding up — a rising API bill, or hitting your Pro/Max limit mid-flow — most of it comes down to a handful of habits. Here are the ten highest-impact ways to spend less, without losing a single capability.

First, the one thing to understand: you're billed for input (everything you send — including the entire conversation so far) and output (what Claude writes back), and output usually costs about 5× input. Crucially, every message in a long chat re-sends the whole history, so long threads get expensive fast. Almost every tactic below follows from that.

1.Start a new chat for each topic

Because the full thread is re-billed on every turn, a long conversation quietly multiplies your cost. When you switch task, open a fresh chat. This single habit fixes more bills than anything else.

2.Ask for the length you want

Output is the expensive part. "Summarise this in 3 bullets" or "answer in one line" can cut an answer's cost by more than half. Don't let Claude pad.

3.Use the right model for the job

Claude comes in tiers. The smallest (Haiku) costs a fraction of the largest (Opus) and handles everyday tasks — summarising, extraction, classification — perfectly. Reserve the top model for genuinely hard reasoning. On a subscription this makes your limit last far longer.

4.Turn off extended thinking for simple tasks

Extended thinking is billed as output. It's brilliant for hard problems and wasteful for "reword this". Leave it off unless the task needs it.

5.Don't make Claude echo your input

"Rewrite this: [long text]" often returns your text and the rewrite — you pay for both. Ask for only the changed version.

6.Summarise and restart long chats

When a useful conversation gets long, ask Claude to summarise the key points, paste that into a fresh chat, and continue. You keep the knowledge and drop the expensive history.

7.Prompt caching — the biggest lever (API)

If you send the same large context repeatedly (a system prompt, a document, examples), you can cache it. Cached reads cost roughly 10% of the normal input price — about 90% off your reused context. For chatbots, agents and RAG this is transformational. Put the static part at the front of your prompt so the cache prefix stays identical.

8.Use the Batch API for bulk work (API)

For anything a human isn't waiting on — enriching data, classifying records, generating descriptions at scale — submit it as a batch for a flat 50% discount, with results within 24 hours.

9.Retrieve, don't dump

Don't paste an entire manual to ask one question. Send only the relevant passages (or cache the manual once). You pay for a few paragraphs instead of a library — and answers get sharper.

10.In Claude Code: /clear, /compact, and point at files

Use /clear between unrelated tasks and /compact when a session grows, so you stop re-billing a giant context. And point Claude at specific files or line ranges rather than dumping a whole repo.

Do just the first three today and most people cut their Claude spend by half or more. For the free quick-reference list, see the token-saving page; for copy-paste prompts that get more from every message, grab the free Prompt Vault.