Why does Claude burn tokens so fast? The 7 real reasons

Claude burns tokens fast because every message re-reads the whole conversation, the newest models always think before they answer, and Fable 5.1 costs far more per token than Opus or Sonnet. Add Claude Code's file reading, cache misses after breaks, agents working in the background and peak hours, and a plan that looked generous empties in an afternoon.

The short answer

How Claude counts what you use

Claude plans do not give you a fixed number of messages. They give you a budget that resets on two clocks: a session window of five hours and a weekly limit. Anthropic's article on understanding usage and length limits lists what moves the needle: conversation length and complexity, message length, the features you use, the model you pick, the effort level, and the files in your projects. It also says something people miss: your usage on claude.ai, Claude Code and Claude Desktop all counts towards the same limit.

So "Claude burns tokens fast" really means one of two things. Either each request is bigger than you think, or more requests are being sent than you think. Every reason below is one or the other. If you want the mechanics of the two clocks first, our guide to how Claude's usage limits really work covers them in detail.

The 7 reasons Claude uses so many tokens

1. Fable 5.1 spends your limit faster than any other model

This is the newest reason and, for many people this autumn, the biggest. Fable 5.1 shipped on 1 September 2026 as Anthropic's most capable generally available model. Anthropic's Help Center is blunt about the cost: Fable models draw from your plan's regular weekly usage limits and use them faster than other Claude models.

Anthropic does not publish how plan usage is weighted per model, but its API prices are the clearest public signal of relative cost, and the gap is wide:

Prices are from Anthropic's launch pages for Fable 5.1, Opus 5.5 and Sonnet 5.5. The same Fable page adds a second multiplier: Fable 5.1 defaults to High effort in Claude Code, and to Medium on claude.ai. So the same task in Claude Code both costs more per token and tends to use more tokens.

How that lands depends on your plan. On Max, Fable can use up to 50% of your weekly limit at no extra cost, then switches to usage credits; we explain that ceiling in Fable's 50% weekly cap on Claude Max. On Pro, Anthropic's pricing page lists Fable under usage credits, so it never comes out of the plan at all; what Claude Pro gets from Fable 5.1 has the details.

The fix: treat Fable as the specialist, not the default. Sonnet 5.5 is, per Anthropic, 30% faster than Sonnet 5 and needs fewer tokens per task. Opus 5.5 got cheaper too, and its launch came with higher five-hour limits on Pro, Max, Team and seat-based Enterprise. Most everyday work does not need the most expensive model in the line-up.

2. Long conversations re-read themselves on every turn

This is the classic one, and it has not gone away. Claude has no memory between requests in the way people imagine. Each time you send a message, the conversation so far goes back to the model with it. Anthropic's Claude Code documentation spells it out: Claude Code sends your full conversation with every request, so a one-line question in a session that has been open all day still draws usage for the whole conversation.

Prompt caching softens this. Content that was just processed can be re-read at a much lower rate, and content stored in a Project is cached and, in Anthropic's words, counts less against your limits when reused. But cheaper is not free. A thread that started with a 40-page PDF is still carrying that PDF on message thirty.

Attachments, Research and web search make each turn heavier still. Anthropic's usage limit best practices name file size, tool use and artifact creation as factors alongside conversation length.

The fix: start a new conversation when the topic changes, and put documents you reuse into a Project rather than re-uploading them. In Claude Code, /clear between unrelated tasks costs nothing. Our list of 13 ways to make a Claude session last longer goes through the rest of these habits.

3. Thinking is always on, and it counts as output

The current top models reason before they reply, and that reasoning is not free. Anthropic's cost guide says thinking tokens are billed as output tokens, and the default budget can be tens of thousands of tokens per request depending on the model. Output is the expensive side of every price in the table above.

On older models you could switch thinking off. On the current ones you cannot: the same guide states that you can't turn off thinking on Opus 5.5, Sonnet 5.5 or the Fable models, which always use extended thinking. What you control is the effort level.

The fix: in Claude Code, lower the effort with /effort or in /model for simple tasks. Renaming a variable does not need the same depth of reasoning as designing a database schema. In the claude.ai app, keep the heavier effort settings for questions that need them.

4. Claude Code explores before it acts

If your limit vanishes faster in Claude Code than in chat, this is why. A chat turn is one request. A Claude Code turn can be a dozen: read a file, search the project, run a command, read the output, edit, run the tests, read the failures. Each tool call sends another request carrying the conversation plus the new results.

Anthropic's guidance for teams budgeting Claude Code puts it plainly: each Claude Code turn carries file contents, tool calls and multi-step reasoning, so one debugging session can consume more than a day of chat. And the size of the job depends heavily on how you ask. The docs contrast "improve this codebase", which triggers broad scanning, with "add input validation to the login function in auth.ts", which lets Claude work with minimal file reads.

The fix: name the file and the function. Use plan mode (Shift+Tab) for bigger changes so Claude proposes an approach before it starts editing, and press Escape early if it heads the wrong way. Keep your CLAUDE.md short, since it loads into every session, and move specialist instructions into skills that load only when needed. Disable MCP servers you are not using.

5. Coming back after a break costs more than you expect

Prompt caching is what keeps long sessions affordable, and it expires. Per Anthropic's cost guide, the cache lifetime is one hour on a subscription and drops to five minutes once you are drawing on usage credits. Your first message after a longer break misses the cache and processes your full context again at the full rate.

That is why a session that felt cheap all morning can take a visible bite out of your limit with a single message after lunch. Compacting has a related cost: /compact reads the whole conversation it summarises, so compacting a large context is itself a large request. When you want a fresh start rather than continuity, /clear is the free option.

The fix: if you are returning to a big session after a long break, decide whether you really need all of that history. On Pro and Max, Claude Code offers to resume from a summary so later requests do not carry the full history. Accept it when the details no longer matter.

6. Work you are not watching is still spending

Some usage happens while your attention is elsewhere. Anthropic's list of reasons usage climbs in a long session includes several of these:

• Subagents. Every subagent sends its own requests on top of the main conversation. They are worth it for keeping noisy output out of your main thread, but they are not free.
• Agent teams. Anthropic's figure is that agent teams use approximately 7x more tokens than standard sessions when teammates run in plan mode, because each teammate keeps its own context window. Each active teammate keeps consuming until it exits.
• Scheduled tasks and loops. A scheduled task fires on its interval even while the session is idle, sending your full context each time.
• Other surfaces. Because claude.ai, Claude Code and Claude Desktop share one limit, a long chat on your phone and an agent run on your laptop drain the same number.

The fix: give helpers a smaller model (Claude Code lets you set model: haiku for a subagent), keep teams small and shut them down when their work is done. On a paid plan, the /usage screen in Claude Code now breaks recent usage down by skills, subagents, plugins and MCP servers, and flags any behaviour that accounts for 10% or more of it. It is the quickest way to find out which of these is yours.

7. Peak hours, and a smaller weekly pool

The last reason is not about how you work but about when, and about changes on Anthropic's side.

Peak hours. In late March 2026, Anthropic confirmed it had made five-hour session limits move faster during weekday peak hours, 5am to 11am Pacific, for Free, Pro and Max, while weekly limits stayed the same. It said about 7% of users would hit session limits they would not have hit before, particularly on Pro. On 6 May 2026, Anthropic removed the peak-hour reduction for Claude Code on Pro and Max, and doubled Claude Code's five-hour limits. That announcement covered Claude Code only. For heavy chat work, the weekday morning window in Pacific time (afternoon in Europe) is still the one to avoid.

A smaller week. If Claude has felt hungrier since mid-September, it is not your imagination. On 14 September 2026 Anthropic ended the temporary 50% weekly boost it had run since May and replaced it with a permanent 25% increase over the old baseline. Anthropic acknowledged the arithmetic: "Compared to today, this works out to a 17% reduction in weekly limits on Claude Code." The same work now takes a bigger share of the week. Our explainer on Claude's weekly limit covers how that pool resets.

The fix: schedule long agent runs and big document jobs outside the peak window if you can, and look at the weekly number, not just the session one. The five-hour session limit resets several times a day; the weekly one does not.

What is usually not the reason

Two explanations come up constantly and rarely hold up. The first is "Claude is broken". Sudden jumps almost always trace back to one of the seven above, most often a long thread, a cache miss or a model switch. The second is "a short message should cost less". As reason 2 shows, the size of a request is set mostly by everything that came before it, not by what you just typed.

One thing that is real but small: Claude Code uses a few tokens in the background even when idle, for summarising past sessions and similar housekeeping. Anthropic puts that at typically under $0.04 per session. It will not empty a plan.

How to see which reason is yours

On claude.ai, Settings then Usage shows your current session and weekly readings, and on Max a separate Fable row. In Claude Code, /usage adds the breakdown by subagent, plugin and MCP server, plus the behaviour flags for long context and cache misses. Between them they answer the question. The catch is that you only see them when you stop and look, which is usually after the expensive part.

That is the gap tokn.watch fills. A remaining percentage in the menu bar that drops 8 points on one message tells you, in the moment, that you just paid for a long context or a cold cache. Connecting takes about a minute, and the setup steps are on our homepage. If you are already out, what to do when you hit your Claude limit covers the next steps, and if you are weighing a second subscription to cover the gaps, should you use Claude or ChatGPT is the honest version of that decision.

The verdict

Claude does not burn tokens for no reason. It burns them because the expensive things are the default things: the newest model is the most expensive, the conversation you are already in is the one you keep typing into, thinking is always on, and Claude Code explores before it edits. None of that is a bug, and all of it is adjustable.

If you change only three habits, make them these: use Fable for the hard problems only, start a new conversation when the topic changes, and give Claude Code specific instructions. Then keep both numbers, session and weekly, where you can see them.

Frequently asked questions