Why is my Claude usage going so fast? What changed in October 2026, what users are reporting, and how to check your own

Why is Claude usage going so fast? If your Claude limit has started running out sooner since mid-October 2026, you aren't the only one. Claude Code users, including some on Max and Team plans, have posted measurements in Anthropic's public GitHub tracker showing their weekly limit draining faster than in early October for similar work. As of October 26, 2026, Anthropic hasn't commented on those reports or announced any October change to limits. A separate open bug report describes Claude Code using a shorter prompt cache than it should, which can make breaks more expensive. Some of the cost has documented causes you can check yourself, starting with the /usage screen. Our Claude app explainer covers how the five-hour and weekly limits work.

This explainer is based on Anthropic's help-center articles on usage limits and limit resets, the Claude Code documentation on costs and prompt caching, Anthropic's Opus 5.5 announcement, the Claude status page, the Claude Code changelog, and issues #103376, #101685 and #104694 in the anthropics/claude-code GitHub repository, checked October 25 and re-checked October 26, 2026. It also draws on reporting by BleepingComputer and The New Stack and on SemiAnalysis's October 5, 2026 newsletter.

The short version

• Users report a drop in mid-October. In issue #103376, opened October 19, a Max 20x user says their weekly meter went from 0% to 54% in 22 hours after the October 18 reset. They say the token volume was similar to the previous week, when about 4.4 days of work moved the meter to 60%. Two other users estimate their weekly limit now covers about 30% fewer tokens than in early October.
• Anthropic hasn't commented. None of the issues has a reply from Anthropic staff, and the status page lists no October incident about usage limits.
• A cache bug report is open. Issue #101685, filed October 13, says interactive sessions write the five-minute cache instead of the one-hour cache that Anthropic's docs say subscribers get. Users report that a one-line setting works around it.
• Some fast usage is by design. Anthropic's docs say long sessions, breaks longer than the cache lifetime, subagents, agent teams, scheduled tasks and compaction all use more of your limit.
• Check before you guess. In Claude Code on a paid plan, /usage breaks down what used your limit and suggests fixes.

What are Claude users reporting in October 2026?

The clearest reports are in issue #103376, "Weekly limit 0% to 54% in 22h with same token volume that took 4.4 days to reach 60% the prior week." The reporter is on Claude Max 20x and uses Claude Code 2.1.299 in the desktop app, with Opus 5.5 as the main model. They counted API calls and tokens from their local transcripts. From October 11 05:00 to October 15 14:22 UTC in the prior week, 5,084 API calls and 1,929.5M cache-read tokens moved the weekly meter from 0% to 60%. From October 18 05:00 to October 19 03:16 UTC, 902 calls and 299.4M cache-read tokens moved it from 0% to 54%.

Last verified: October 26, 2026

Two other users added data to the same issue: one on a Max account, and one on a Team seat using Claude Code 2.1.304. Both weighted their tokens by API price and divided by how far the weekly meter moved. Each estimated that 1% of the weekly meter now covers about 30% fewer weighted tokens than it did in early October. The Team user said their total weekly volume wasn't higher than in September, but they hit 100% of the weekly limit on October 8 and again on October 16.

The three sets of numbers don't fully agree. The original reporter's post-reset drop is far steeper than the other two users saw, and they suggested it might be two effects: a gradual reduction, plus something that hit some accounts after the October 18 reset. Several more users, including Max 20x subscribers, have since said in the thread that they see the same thing, without posting measurements, and on October 23 another Max user opened a separate issue, #104694, saying 30% of their weekly quota went in less than 24 hours. These are reports by individual users. They aren't verified, and as of October 26 both issues are open with no reply from Anthropic.

Is there a Claude usage bug right now?

There's an open bug report, but Anthropic hasn't confirmed a bug. Issue #101685, filed October 13, says interactive Claude Code sessions on Team and Max plans write their prompt cache with a five-minute lifetime, even though the client asks for one hour. The report covers Claude Code 2.1.296 to 2.1.300, and later commenters reported the same behavior through 2.1.304. Reporters put the start at different times between about October 12 and October 19, varying by account and session, and they say claude -p runs on the same accounts still wrote one-hour caches.

That matters because of how Anthropic says caching works. In a long session, each message re-sends your whole conversation. With a warm cache, most of it is billed at the cheap cached-read rate. The Claude Code docs say that on a subscription within your plan's usage, the main conversation gets a one-hour cache by default. After a break longer than the cache lifetime, the next message reprocesses everything at full cost. On a five-minute cache, even a coffee break triggers that. Anthropic's docs also say cache reads are billed below the standard input rate, while one-hour cache writes are charged at a higher rate than five-minute ones, as our explainer on Claude Code's idle compaction details.

Users in the thread disagree on how much it costs in practice, and their idle tests conflict. One commenter found that caches labelled five-minute still survived a 30-minute idle gap and had expired by 70 minutes, which suggests the label may not match the real lifetime. The original reporter's single test saw the conversation written to the cache again after 11.5 minutes idle, and another user saw a full cache read after a 9.5-minute gap. The reporter traced the five-minute writes to a beta flag, cache-keepalive-2026-09-03, that interactive sessions send. Anthropic hasn't said whether this is intended, and none of the Claude Code releases up to 2.1.310, published October 23, lists a fix for it in the changelog.

The workaround users report is Anthropic's own documented setting. Add this to ~/.claude/settings.json (Claude Code v2.1.242 or later):

{ "promptCacheTtl": "1h" }

To see which cache your requests use, Anthropic's docs suggest running claude -p "hello" --output-format json. One-hour writes appear under ephemeral_1h_input_tokens and five-minute writes under ephemeral_5m_input_tokens. That command runs a -p session, though, and the bug report says -p wasn't affected. To check interactive sessions, the reporters read the same fields from their local transcripts.

Did Anthropic cut Claude's limits?

There's no announcement of a cut in October. The last announced changes were:

On September 14, Claude Code's weekly limits dropped back from a temporary 50% boost. Anthropic said the new permanent level was 25% above the original baseline. In a clarification BleepingComputer reported on August 29, it added: "Compared to today, this works out to a 17% reduction in weekly limits on Claude Code." If your weekly limit felt tighter in mid-September, that's the documented reason. It doesn't explain a change between early and late October.

On September 22, Anthropic's Opus 5.5 announcement said it was "increasing five-hour usage limits" on paid plans. It didn't mention weekly limits. The same day, Anthropic gave subscribers a limit reset they could save and use later. Its help article says an unused reset "expires on the day and time listed on the offer," shown in Settings > Usage. If you used that reset, or let it expire, a later week can feel tighter by comparison.

Anthropic can also run limit experiments on some accounts. In its October 5 newsletter, SemiAnalysis said that one of three subscriptions to the same plan it tested at one provider had about 20% lower limits. The provider told SemiAnalysis this was an "extremely tiny" A/B test of "how to better balance when people hit limits." The part of the newsletter we could read doesn't name the provider.

This has happened before. In late March 2026, Anthropic said on Reddit, "We're aware people are hitting usage limits in Claude Code way faster than expected," and called it the team's top priority, The New Stack reported. Anthropic's Lydia Hallie later said the team had shipped fixes on the Claude Code side, according to The AI Economy.

What uses up Claude usage so fast?

Even without a bug, some patterns use far more of your limit than they seem to. Anthropic's Claude Code cost docs list these causes of rising usage in a long session:

• Long context. Every request re-sends the whole conversation, and every tool call sends it again with the results. A one-line question in a session that has been open all day still pays for the full history.
• Cache misses. Your first message after a break longer than the cache lifetime reprocesses everything at full cost. Anthropic's prompt-caching docs add that switching models, turning on fast mode and, when tool search is off, connecting an MCP server mid-session also miss the cache.
• Scheduled tasks. /loop and other session-scoped tasks keep firing while you're away, each one sending the full context.
• Messages from other sessions and goal check-ins. These start new turns while the session sits idle. Setting crossSessionInbound to hold holds messages from other sessions, and setting CLAUDE_CODE_GOAL_CHECKIN_MINUTES=0 turns check-ins off.
• Subagents, workflows and agent teams. Each one sends its own requests. Subagents use a five-minute cache even on a subscription unless you change it, and Anthropic says agent teams use about 7× the tokens of a standard session when teammates run in plan mode.
• Compaction. /compact reads the whole conversation to summarize it. /clear costs nothing.
• Model and effort. Anthropic's help center lists the model and your effort level among the factors that affect usage. The limits are shared across claude.ai, Claude Code and Claude Desktop.

Claude Code's auto mode no longer counts its classifier calls against your limit, as our auto mode guide explains, so that isn't a cause.

How do I check what's using my Claude usage?

Start with Claude Code's own tools:

1. Run /usage. On Pro, Max, Team and Enterprise plans it shows your plan usage bars. It also shows a breakdown of recent usage by skill, subagent, plugin and MCP server, with a recommendation for anything behind 10% or more. Press d or w to switch between the last 24 hours and the last 7 days. The breakdown covers this machine only, not claude.ai or other devices.
2. Read the "Prompt cache (main)" line in /usage (v2.1.251 or later; the likely-cause text needs v2.1.260). It shows your cache hit rate, misses, the likely cause of the last miss and whether the cache is on a one-hour or five-minute lifetime. If it says five minutes and you aren't using usage credits, you may be affected by the bug report above.
3. Press f in /usage (v2.1.305 or later) to have Claude Code apply the recommendations it can.
4. Check the claude.ai usage page (Settings > Usage). Anthropic's help center says it shows any limit reset you have and when it expires, and the reporter in issue #103376 used it to see how their usage split between chats, Claude Code and Cowork.

If your numbers still don't add up, Anthropic's docs say to contact support through Get help under your initials on claude.ai. For a reproducible metering problem, the GitHub issues above are where other users are comparing data. Our guide to the free plan covers what happens when you hit a limit on each plan.

What's confirmed, reported and uncertain

• Confirmed (Anthropic):
  • the five-hour session limit and weekly limit, and the factors that affect usage;
  • the one-hour cache default for subscribers' main conversation, and the promptCacheTtl setting;
  • the causes of fast usage and the /usage tools in the Claude Code docs;
  • the five-hour limit increase and saved reset with Opus 5.5 on September 22;
  • no October status-page incident about usage limits.
• Reported:
  • the September 14 change and the 17% figure, from Anthropic's post as reported by BleepingComputer;
  • the faster weekly meter in mid-October, from user measurements in issue #103376 and a user report in issue #104694;
  • the five-minute cache writes in interactive sessions, from user reports in issue #101685;
  • the A/B test at an unnamed provider, from SemiAnalysis.
• Uncertain:
  • whether Anthropic changed weekly limits, metering or prices in October;
  • whether the cache report is a bug, and how much it costs users;
  • whether your account is in any test. Anthropic hasn't commented on the GitHub issues.