If your Claude Code usage suddenly dropped faster than usual, the cause is usually one of four things: a long conversation piling up context, a cache reset after stepping away, extended thinking (the model's internal reasoning) running by default, or repeated tool calls and MCP servers filling every turn with overhead. Working through the checks below usually takes one to three minutes and points straight at the fix.
🟢 Model references match the current lineup · model notice · Fable subscription
Fable 5 and 5.1 subscription (updated September 7, 2026): Claude Fable 5.1, released September 1, 2026, is the current Fable model and Fable 5 is now legacy. Plan terms are the same for both — Max and Team Premium plans include Fable at up to 50% of the weekly usage limit; Pro and Team Standard use usage credits (
Before you start
No special setup is needed for the steps below. As long as Claude Code is installed and you're signed in, you can follow along.
- Claude Code installed and signed in — you need a terminal where you can type commands.
- Know your plan — Pro, Max, Team, and Enterprise usage is managed by a rolling 5-hour session limit and a weekly limit, while API/cloud-provider usage is billed per token with no such limit.
- Check your version — screens can differ slightly by version, so if what you see doesn't match this guide, trust what's actually on your screen.
Terms worth knowing first
Six terms come up repeatedly below, so it helps to define them up front.
| Term | Plain-language meaning |
|---|---|
| Context window | The space that holds your conversation and file contents so far. The fuller it gets, the heavier every request becomes. |
| Prompt caching | A feature that briefly stores the repeated early part of a conversation on the server so re-reading it costs less. |
| Extended thinking | Internal reasoning Claude does before answering. That reasoning is billed as output tokens. |
| Subagent | A helper agent that works in its own separate context on one task and returns only a summary to the main conversation. |
| MCP (Model Context Protocol) | The protocol Claude uses to connect to external tools and services. More connected servers means more overhead added to every turn. |
| Compact | Summarizing and shrinking a long conversation, run with the /compact command. |
Why does usage climb as a conversation gets longer?
Because Claude Code resends your entire conversation with every request. Each time a tool is used, another request goes out carrying that tool's results, so a session left open for hours can bill a short question at the cost of the whole accumulated conversation. Prompt caching processes the repeated earlier portion at a cheaper cached rate, but it doesn't change the underlying fact that the conversation itself is long.
Why does usage suddenly jump after a short break?
Because your prompt cache expired, so the full conversation gets reprocessed from scratch. On a subscription, that cache lasts an hour, but it drops to five minutes while you're drawing on usage credits past your plan limit; on an API key or cloud provider, it's five minutes by default. So if a subscriber steps away for a lunch break longer than an hour and comes back with any question, that single request has to re-read the entire backlog of conversation, consuming far more usage than usual in one shot. On Pro and Max, resuming a large session after a long break sometimes prompts you to continue from a summary instead — accepting that means later requests carry the summary rather than the full history.
Why does extended thinking increase usage?
Because the reasoning it produces is billed as output tokens. Claude Code turns extended thinking on by default to improve quality on complex planning and reasoning tasks, and depending on the model, the default thinking budget for a single request can run to tens of thousands of tokens. Letting every simple task think that deeply adds up in output tokens. Note that Fable-family models always use extended thinking and it can't be turned off there.
Why do repeated tool calls or multiple MCP servers increase usage?
Because each MCP server's tool listings and instructions ride along in your context on every turn. Tool definitions are deferred by default, though, so only a tool's name and its server's instructions enter context until you actually use that tool. Even so, having many servers connected — especially ones you rarely use — adds up. CLI tools like gh, aws, and gcloud carry no such per-tool listing, so the same task runs lighter through them.
How to check where your usage is actually going
Run the /usage command. /cost and /stats are aliases that open the same screen.
- In your Claude Code session, type
/usageand run it. - Check the Session block for this session's token usage and estimated cost. On Pro, Max, Team, and Enterprise, it also shows what share of recent usage went to skills, subagents, plugins, and individual MCP servers.
- This means it worked — if a factor like long context or cache misses accounts for 10% or more of recent usage, it's flagged separately near the bottom. Start your cuts with whatever gets flagged there for the biggest impact.
Ways to reduce usage
Each of the eight items below is a single command or a one-time setting change.
| Method | How to do it | Sign it worked |
|---|---|---|
| Move to a completely new topic | Run /clear | The session cost figure restarts from zero. |
| Keep context while trimming it | Run /compact, or add instructions like /compact keep only code samples and API usage | The conversation is summarized and you can keep working. |
| Use a lighter model for simple work | Switch to Haiku or Sonnet with /model; set model: haiku in a subagent's config | Responses come back faster and the cost climbs more slowly. |
| Lower reasoning depth when deep thinking isn't needed | Adjust with /effort, or turn off extended thinking in /config (on fixed-budget models, lower the MAX_THINKING_TOKENS environment variable) | Answers arrive faster and use fewer output tokens. |
| Clean up unused MCP servers | Check the list with /mcp and disable ones you don't use | Less tool-listing overhead rides along on every turn. |
| See what's filling your context right now | Run /context | You get a breakdown of what's taking up space. |
| Offload long logs or test output to a subagent | Delegate that task to a subagent | The verbose output stays with the subagent; the main conversation gets only a summary. |
| Keep CLAUDE.md lean | Stay under 200 lines; move detailed workflows into skills | Less baseline content is loaded into every turn. |
When should you use /clear versus /compact?
Use /clear when you're switching to a completely different topic — it wipes the conversation without reading it, so it costs nothing. Use /compact when you want to keep working on the same task but need to free up space; the catch is that building the summary requires re-reading the entire conversation once, so that single summarization request itself is expensive. In short: starting fresh calls for /clear, and preserving context is worth the cost of /compact.
Common sticking points and fixes
Three sticking points come up most often.
- Running
/compactin a fresh session prints "Not enough messages to compact." — Cause: there's no conversation history yet to summarize. Fix: proceed with the conversation for a bit before trying again, or if it's already short there's nothing to trim. - A "You've hit your session limit" or "You've hit your weekly limit" message appears — Cause: you've reached your subscription plan's rolling 5-hour or weekly limit, which is shared across all models. Fix: wait for the reset time shown on screen, or increase it with
/usage-creditsif usage credits are turned on for you. For the full breakdown of plan limits, see the usage limits guide. - There's no option to turn off extended thinking — Cause: if you're on a Fable-family model, it always uses extended thinking and that can't be disabled. Fix: switch to a different model if you want to turn thinking off.
Taking it further
Once you've applied the above, two more things are worth trying. First, run /insights to get a report on your recent sessions so you can see exactly which kinds of work drive your usage up. Second, configure your status line to show context usage continuously, so you don't need to type /context every time to see how full the conversation is. For the exact limit numbers on each plan, check the usage limits guide.
Frequently asked questions
Q. Does leaving a session open with no activity still use tokens?
Yes, a small amount. Background jobs that summarize past conversations, along with processing for some commands, use tokens even without interaction, typically under $0.04 per session. If scheduled tasks or cross-session messages are connected, though, each one resends the full conversation and can add up to much more.
Q. Does switching from Opus to Sonnet or Haiku reliably cut usage?
Yes, switching to a lighter model is one of the most reliable ways to cut usage. Sonnet handles everyday coding tasks well, so reserve Opus for cases that genuinely need complex architectural judgment, and set simple repetitive subagent tasks to use Haiku for an even bigger gap.
Q. How much does usage grow when running multiple subagents or an agent team at once?
Usage scales roughly with the number of active subagents or teammates, since each one keeps its own separate context window. Running several agents in plan mode at the same time in particular can use far more tokens than usual, so keep teams small and shut down teammates as soon as their work is done.
Related articles
- When Claude Refuses to Translate: How to Diagnose and Fix It
- Is Claude's Usage Draining Too Fast? How Bug Fixes and Limit Resets Actually Work
- Claude Opus 5 Suddenly Feels Off? 3 Real Reasons, Explained
- Claude Sonnet 5's API Price Hike Was Cancelled — Here's What Changed
- No Design Skills? Here's How to Finish Your UI with Claude Code