13 Ways to Save Claude Tokens and Make Your Session Last Longer
The fastest way to save Claude tokens in 2026 is to stop using Opus for everything (about a 5x reduction on its own), start new chats more often (token cost grows quadratically with conversation length), and run /compact when your context gauge crosses ~50%. Stack ten more boring tactics on top of those three and you can typically extend a Pro user's weekly cap by 30-50% with no change in the actual work you ship. Here are the 13 that move the needle, sourced from Anthropic's docs and the heavy-user community, with the reason each one works.
Why this matters more on Pro and Max than it used to
Until mid-2025, Claude was capped per session - hit the wall, wait five hours, keep working. That single-window model rewarded short bursts of intense use. The weekly cap that Anthropic added on top changed the math: it is now possible to burn your entire week on a Tuesday afternoon by mishandling context, and there is no fix until your 7-day rolling window resets. For the underlying mechanics see how Claude's usage limits really work in 2026 - this article is the tactical companion piece.
Saving tokens does not mean producing less. It means producing the same work without sending Claude redundant context, without leaving Opus running on tasks Sonnet handles fine, and without retrying tool calls because the cluster was overloaded. The 13 tactics below are ordered roughly by impact. The first five do most of the work; the last eight are diminishing-return optimisations that compound. If you first want to know where the usage goes, we break down why Claude burns tokens so fast.
The 13 tactics
Default to Sonnet. Use Opus on purpose.
Estimated impact: 30-50% of total savingsOpus generates substantially more internal reasoning tokens per turn than Sonnet on the same prompt, even for tasks Sonnet would handle correctly. The exact ratio varies by task but roughly 3-5x is the typical observation for code-heavy work. Anthropic's own model selection guide recommends Sonnet as the default and reserves Opus for tasks that genuinely need deeper reasoning.
The mistake is leaving the model picker on Opus "just in case". Set Sonnet as your default in claude.ai (Settings → Model) and in Claude Code (it is already the default there). Switch to Opus deliberately for architectural design, novel debugging, or dense legal/scientific text - then switch back. If you are not actively thinking "this task needs Opus", you are wasting tokens.
Start new chats more often than you think you should.
Estimated impact: 15-25%The cost of one Claude turn scales with the entire conversation history. Turn 50 of a long chat re-sends turns 1 through 49 as context, every single time. A 50-turn conversation about three unrelated topics costs roughly 5x what three separate 17-turn conversations would cost - same output, same number of turns, dramatically different bill against your weekly cap.
The trigger to start a new chat is simple: the topic genuinely changed. Switching from "help me refactor this Swift file" to "draft a follow-up email"? New chat. Coming back tomorrow on a different feature? New chat. The mental cost of losing context is real but small - most context that mattered for the previous task is irrelevant to the next.
Run /compact when your Claude Code context gauge crosses ~50%.
Estimated impact: 10-20% on long sessionsThe /compact slash command in Claude Code asks Claude to summarise the current conversation in place and replace the long history with a short, dense summary. A chat that was 60% through its context window can drop to 15% after a /compact, freeing up the rest of the session for actual work. The summary preserves the high-level state (what you are building, what is done) and drops the verbose tool-call output, which is where most of the bloat lives.
Run it pre-emptively: when you see the context gauge crossing 50%, before you start a new sub-task, and especially before any "do something big" turn that will generate a lot of tool calls. The compacted context is enough to continue working; the raw history rarely is.
Use /clear between unrelated tasks in the same session.
Estimated impact: 5-15%/clear is /compact's stricter sibling: it drops the conversation history entirely and starts fresh, while keeping the same Claude Code session alive. Use it when you are switching to a genuinely different task and none of the prior context matters - debugging a different file, drafting a different feature, hopping into a different repo. Faster and cheaper than a /compact because there is nothing to summarise.
The rule of thumb: /compact when you want to keep working on the same thing but the history is bloated; /clear when the next task is unrelated to the previous one. Both are dramatically cheaper than letting context fill up to 100% and watching Claude either refuse to continue or auto-truncate in unhelpful ways.
Prefer Edit over Write for files that already exist.
Estimated impact: 5-10%Claude Code's Edit tool sends only the diff - the exact substring you are replacing and the new substring. The Write tool re-sends the entire file body, even for a one-line change. On a 1,000-line file, that is a 100x difference in tokens for the same edit. Anthropic's tool reference for Claude Code calls this out explicitly as the preferred path for modifying existing files.
You do not have to tell Claude this directly every turn - the system prompt already prefers Edit for existing files. The mistake is asking for "rewrite this file" when "edit this function" would do. Phrasing matters: "rewrite" pushes toward Write; "edit" or "change" pushes toward Edit.
Tell Claude what to grep for, not "find the right file".
Estimated impact: 5-10%"Scan the codebase for X" can turn into ten Read calls before Claude lands on the right file. "Grep for the symbol foo in src/, then Read just the matching file" turns into one Grep + one Read. The difference is 8-9 wasted Reads, each one expensive on a non-trivial file.
The general principle: search before you read. Claude knows how to use Grep - it just needs the explicit instruction to use it first. A short prompt like "grep for X to find the relevant file, then read it" replaces a sprawling exploration loop with a targeted lookup. Saves tokens, saves time, gives you a more focused answer.
Spawn subagents for parallelizable work that does not need to live in your main context.
Estimated impact: 5-15% on agentic workflowsThe Task tool in Claude Code spawns a subagent with its own context window. The subagent does the work - reading files, grepping, exploring, running checks - and only the final summary returns to the parent context. The exploration noise stays in the subagent, where it is cheaper and shorter-lived.
The natural fit: "explore this codebase and tell me where authentication is handled" or "find all the places that call this function and check their error handling". These are bounded exploration jobs with a clear deliverable; they are the worst possible thing to do in main context and the best possible thing to delegate. Use subagents whenever the task is "go figure something out and report back".
Route classification and triage subtasks to Haiku via the API.
Estimated impact: 2-8% on heavy workflowsIf you build pipelines on top of the Claude API, Haiku is roughly 3x cheaper than Sonnet and 5x cheaper than Opus on current per-token pricing. For jobs like "which of these 50 files is most likely relevant" or "classify this PR as feature/bugfix/refactor/docs", Haiku is fast, cheap, and accurate enough. Anthropic publishes the per-token pricing in the API overview.
This one does not apply if you use Pro or Max exclusively through the web/Claude Code, since those are flat-fee subscriptions. But anyone building agentic workflows where some steps need careful reasoning and other steps just need triage should split the load: Sonnet for the thinking, Haiku for the routing. Cuts the API bill significantly and tightens the subscription cap if you mix both surfaces.
Keep a tight CLAUDE.md at your repo root.
Estimated impact: 3-8% (saves opening cost per session)Claude Code automatically loads a CLAUDE.md from the repo root into every session in that project. Spending 30 lines explaining "this is a Swift app, entry point is X, conventions are Y, don't touch Z" once saves the first 2-5 turns of every subsequent session. Per Anthropic's memory and project files docs, this is the supported pattern.
Keep it dense: 50-200 lines is the sweet spot. A 2,000-line CLAUDE.md is the opposite of a token-saver - it loads into context on every turn whether you need that section or not. Sections that go stale (specific TODOs, in-flight feature notes) should live elsewhere; CLAUDE.md is for facts about the project that change slowly.
Turn off extended thinking for routine edits.
Estimated impact: 2-10%Extended thinking adds hidden reasoning tokens to each turn - useful for hard problems, wasteful on quick edits and grep-like work. On both claude.ai and Claude Code, thinking can be toggled on a per-turn basis or set as a session default. The cost difference per turn is real and adds up over a long working session.
Heuristic: if you are asking Claude to do something you could mostly do yourself but slower, thinking is overkill. If you are asking Claude to do something where the answer is not obvious to you, thinking helps. Default to off for routine code edits, file moves, and structured-format conversions. Default to on for debugging, design, and dense explanations.
Externalise the plan to TodoWrite instead of re-explaining it each turn.
Estimated impact: 2-5% on long agentic runsOn long Claude Code sessions, it is tempting to start each turn with "OK so the plan is X, Y, Z, we did X, now doing Y". Each re-statement costs tokens. The TodoWrite tool stores the plan in a structured side channel; Claude can update one entry per turn instead of re-narrating the whole list. Net effect: less filler in every turn, more budget for actual work.
This is a habit-change, not a setting. Once the plan lives in TodoWrite, Claude generally references it implicitly ("marking step 2 done, starting step 3") instead of restating it. Adopt the habit early in any multi-step task; it pays off immediately on jobs of more than ~5 steps.
Heavy work outside US working hours (roughly 9 AM-5 PM PT).
Estimated impact: 2-5% (avoided retries)Anthropic's infrastructure is most loaded during US working hours. On peak hours, Pro users see more "Claude is at capacity, try again" errors, more failed tool calls that have to be retried, and slower model responses overall. The retried tool calls cost you tokens twice for the same work - a 5% baseline retry rate becomes 5% wasted budget you cannot recover.
If your work is asynchronous, batch the heavy stuff into off-peak windows. For European users this is the natural afternoon/evening already. For US users it means kicking off long Claude Code agentic runs early morning or late evening. The model is the same; the cluster headroom is not.
Be specific about what you want; avoid open-ended exploration prompts.
Estimated impact: 3-10% (varies by prompt style)"Look around this codebase and tell me what you think" is the most expensive prompt you can write - Claude will read widely, write a lot, and generally chase its tail. "Compare the implementations of X and Y in files A.swift and B.swift and tell me which one handles the edge case Z" is the cheapest prompt that does useful work. Both can produce the same insight; the second costs a fraction of the tokens.
The discipline is the same one that makes you a better tech-lead: scope tightly, ask for the deliverable shape, name the files. Claude is good at execution and weak at framing. You bring the framing. The token savings are a side effect; the better answers are the main reason.
Install an early-warning meter so you see the cap coming.
None of these tactics matter if you only notice the weekly cap at 0%. tokn.watch shows your 5-hour, weekly, and Opus percentages live in your menu bar - free, no account, reads your existing browser session. You see the percentage falling and pace yourself before it bites, instead of explaining away the surprise on Wednesday afternoon.
DOWNLOAD FREEWhat the impact actually looks like, stacked
The table below is the rough cumulative effect of layering the first ten tactics on top of a baseline "I use Claude Code on Opus, never compact, never start new chats" user. Numbers are estimates derived from heavy-user benchmarks shared on Hacker News and Reddit threads about Claude usage; treat them as directional, not exact.
| Tactic | Cumulative weekly-cap headroom gained |
|---|---|
| Baseline (Opus always, no /compact, long chats) | 0% (you are here) |
| + Sonnet default | ~30-50% |
| + Start new chats on topic change | ~50-65% |
| + /compact at 50% gauge | ~60-75% |
| + /clear between unrelated tasks | ~65-80% |
| + Edit over Write | ~70-85% |
| + Grep before Read + subagents + CLAUDE.md | ~80-100% |
The reason the early tactics dominate is straightforward: token cost is asymmetric. Opus-everywhere is expensive; Sonnet handles 90% of work. Long chats are expensive; new chats cost the same as the first turn. Once those two are fixed, the remaining optimisations are real but smaller. Do not skip the small ones - they compound - but spend your habit-changing energy on the top two first.
The tactics that look smart but do not actually save tokens
A few common pieces of advice that get repeated in Claude communities do not survive measurement. Worth naming them so you do not waste time on them.
- "Use shorter prompts." Prompt length is a small fraction of total turn cost on anything but trivial requests. The response and tool calls dominate. A 200-token prompt that gets you a focused answer is cheaper than a 50-token prompt that triggers ten exploratory tool calls.
- "Compress your code before pasting." Minified or whitespace-stripped code is harder for Claude to reason about, leading to more back-and-forth turns that wipe out the saving. Paste readable code or, better, give Claude a file path.
- "Avoid markdown formatting." Markdown tokens are negligible. Code blocks, headers, and bullet points are essentially free and help Claude parse your intent.
- "Use a smaller font." Yes, this advice exists. Font has no relationship to tokens.
The general rule: if a tactic sounds clever but does not change what Claude reads, generates, or how often, it is unlikely to save tokens.
How to verify these tactics actually work for you
Estimates in this article are rough. The honest way to evaluate any tactic is to measure, change one habit at a time, and re-measure. Two-week sprints work well:
- Baseline week: Use Claude however you normally do. Note your weekly all-models percentage every Friday afternoon. That is your starting point.
- Change one habit: Pick the highest-impact tactic you are not currently doing - usually Sonnet-default or starting new chats more often. Do that one consistently for two weeks.
- Re-measure: Check the percentage again. If you have more headroom, the tactic works for your workflow. If not, drop it and try the next one.
- Stack: Once you have one habit locked in, add the next one. Compounds quickly.
The instrument you need for this is the weekly percentage. Anthropic surfaces it on Settings > Usage on claude.ai (manual checks), via /usage in Claude Code (on-demand), and in third-party menu-bar trackers (continuous). For the full comparison see the complete guide to tracking your Claude usage - the four approaches all read the same number, they just trade freshness for friction.
If you are not sure whether you are even hitting the cap, the gating question is upstream of this article: see Claude Pro vs Max - which one matches your usage for the decision framework. Most readers who land here turn out to be Pro users with room to spare; the tactics in this piece are the right move for them. The smaller group who are genuinely capped out should be on Max 5x already and are using these tactics to defer the next upgrade.
Frequently asked questions
How do I save Claude tokens?
The five highest-impact tactics in 2026: default to Sonnet instead of Opus (Opus uses about 5x more tokens per turn); start new chats more often (token cost grows quadratically with conversation length); run /compact in Claude Code when your context gauge hits ~50%; prefer Edit over Write for existing files (Edit sends only the diff); and use Grep before Read so Claude finds the file before reading it. Combined, those five typically extend a Pro user's weekly cap by 30-50% with no change in actual output.
Does Claude charge per token on Pro and Max?
Pro and Max are flat-fee subscriptions, not per-token billing. But the plans do enforce token-based usage caps (the 5-hour session window and the 7-day weekly window), so saving tokens directly extends how long your subscription lasts before you hit the cap. The API is the only Claude product that bills per-token at consumption time. If you are on a Pro or Max subscription and never hit the weekly cap, you can stop reading and pocket the optimisation.
What is /compact in Claude Code?
/compact is a Claude Code slash command that summarises the current conversation in place and replaces the long history with a short, dense summary - then continues where it left off. The effect is dramatic on long sessions: a chat that was 60% through its context window can drop to 15% after a /compact, freeing up the rest of the session for real work. Run it whenever the context gauge crosses ~50%, especially before kicking off a new sub-task in the same session.
When should I start a new Claude chat versus continuing?
Whenever the topic genuinely changes. The cost of one Claude turn scales with the entire conversation history - turn 50 of a long chat costs roughly 50x what turn 1 cost in token expense. If you are switching from "help me refactor this file" to "help me draft a marketing email", start a new chat. The mental cost of losing context is real but small; the token cost of carrying irrelevant context forward is enormous.
Should I avoid using Opus to save tokens?
Default to Sonnet, reserve Opus for tasks that genuinely need it. Opus is roughly 5x the per-turn token expense of Sonnet on the same prompt because it generates more tokens internally during reasoning. For routine coding, refactors, code review, drafting, and most chat use, Sonnet handles it at a fraction of the cost. Opus is the right choice for hard reasoning - architectural design, novel debugging, dense legal or scientific text. The mistake is leaving the model picker on Opus for everything "just in case".
What is the peak-hour rule for Claude?
Anthropic's infrastructure is most loaded during US working hours, roughly 9 AM-5 PM Pacific. Pro users on peak hours see more "Claude is at capacity, try again" errors, slower responses, and more failed tool calls that have to be retried (which costs you tokens twice for the same task). Running the same work at 6 AM PT or 9 PM PT typically gives smoother responses for the same token spend. For European users this is the natural afternoon/evening window already; for US users it means batching heavy Claude Code work outside the working day where possible.
Does a CLAUDE.md file save tokens?
Yes, indirectly. A short CLAUDE.md at your repo root loads automatically into every Claude Code session in that project, so you do not have to spend the first few turns of every session explaining "this is a Swift app, the entry point is X, here are the conventions". Each of those re-explanation turns costs tokens; the CLAUDE.md costs once and saves every session afterwards. Keep it dense - 50-200 lines is the sweet spot. A 2,000-line CLAUDE.md is the opposite of a token-saver.
Will I notice the difference if I use these tactics?
Yes, but only if you measure. The Settings > Usage page on claude.ai shows your weekly remaining percentage, and an ambient tracker like tokn.watch shows it live in your menu bar. The honest way to evaluate any of these tactics is to install the tracker, note where your percentage usually lands on Friday, change one habit for two weeks, and check Friday again. Without the measurement, you will overestimate the impact of fashionable tactics and underestimate the boring ones (Sonnet-default and Edit-over-Write are the two highest-impact tactics and the two most boring).
See your Claude cap falling in real time, right in your menu bar.
Free macOS app. Reads your existing browser session - no account, no servers, no API key. Shows your Claude 5-hour, weekly, and Opus percentages live, alongside ChatGPT if you use both.
DOWNLOAD FREELast verified: 2026-07-21. We re-check the facts in this article every 90 days. Anthropic occasionally renames slash commands and adjusts model defaults; the underlying tactics (model picking, context hygiene, compaction, peak-hour avoidance, externalising plan state) have been stable since the weekly cap launched.
Sources: Anthropic - choosing a model; Anthropic - Claude Code quickstart; Anthropic - Claude Code memory and project files; Anthropic - API overview and pricing; Anthropic - about Claude Pro usage; Anthropic - prompt caching.