How to Save Claude Tokens

Everything Claude reads and writes costs tokens. That covers your full conversation history, files, tool results, and instructions, not only your newest message.

Saving tokens gets you:

• Fewer hits on usage limits
• Faster answers
• Smaller API bills
• Sharper focus, since Claude reads less noise

Long, cluttered chats are the largest hidden cost.

Every Claude plan runs on a rolling five-hour usage window, and paid plans add weekly limits on top. Once a paid plan hits a limit, it can buy extra usage credits at API rates.

Larger models, higher effort, long chats, big files, and tools all burn through your limit faster.

Haiku

Fastest and cheapest. Simple tasks, summaries, and extraction.

Sonnet 5.5

Most day-to-day work and coding.

Opus 5.5

Hard reasoning and complex projects.

Fable 5.1

Long-running agents and the toughest problems. Drains limits fastest.

Begin with a smaller model and step up only when the quality demands it.

Claude’s effort levels run from low to max. More effort means more thinking tokens.

• Low or medium for routine tasks
• High for planning and debugging
• Max only for the toughest problems

In Claude Code, adjust it with /effort.

• Open a new chat for every new topic
• Request a short summary before switching chats
• Paste only that summary into the new chat
• Never keep a single chat running for days

In a long chat, each new message re-reads everything that came before.

When an answer misses, edit and resend your original message rather than adding a correction. That swaps out the wrong turn instead of piling another on top.

• Give Claude the context it needs, once
• Bundle related questions into one message
• Specify the length and format you want
• Request bullet points or tables over long prose
• Say “no preamble” when you just need the answer

• Share only the pages or sections that matter
• Convert huge PDFs to plain text when you can
• Strip out repeated boilerplate
• Keep reused files in Projects instead of re-uploading them

Each connector, tool, and plugin adds instructions Claude has to read. Switch off connectors and web search for tasks that do not use them.

Claude Code loads CLAUDE.md at the start of each session. Keep it brief:

• Only the rules Claude truly needs
• No long histories or pasted docs
• Links to files instead of copied content
• No rules that have stopped applying

/clear

Start clean when you switch tasks

/compact

Condense a long session to free up context

/model and /effort

Drop to cheaper settings for simple work

/mcp

See which servers are connected, and disconnect the ones you do not use

Point Claude at specific files rather than asking it to explore the entire codebase.

Subagents read and search inside their own context, which keeps your main session lean. Run them on a cheaper model, like Haiku, for simple research and file reading.

• Cache prompts that repeat instructions and documents
• Send non-urgent jobs through the Batch API, at 50% off
• Pick Haiku 5.5 for high-volume tasks
• Keep Haiku 5.5 prompts below 100K tokens to avoid higher rates
• Cap maximum output length
• Count tokens before sending large requests

• Plan big work in a single message, not many small ones
• Tackle heavy tasks early in your usage window
• Save strong prompts as skills instead of retyping them
• Request a plan before long tasks
• Stop a task early when it heads the wrong way

Answer concisely. No preamble, no repetition. Use bullet points and keep the answer under [length]. If you need more information, ask one short question rather than guessing. Task: [task].