Claude Pro Limit Exhausted After Every 5-Hour Reset? Here's the Fix

Key Takeaways

• Run claude --kill-all to stop background sessions draining your Claude Pro quota.
• Check claude.ai/usage for stuck processes.
• If the issue persists, contact Anthropic support.

The Problem: Quota Drained Before You Even Type

A Claude Pro user on Reddit reports a frustrating issue: after hitting the 5-hour usage limit, they wait for the reset, only to see "Usage limit reached" the moment it resets. They can't send a single message, yet their weekly usage keeps climbing. They've tried logging out, powering off, and deleting conversations—nothing works.

This isn't a normal "Claude Code eats tokens" complaint. The user is doing simple prompt→edit workflows, using /compact at 60% context, and has no agents or background tasks. Yet the quota is consumed before they even act.

The Likely Culprit: Background Sessions

Claude Code can leave background processes running, especially if you've used features like background tasks, hooks, or MCP servers. These processes may keep making API calls even after you close the terminal. The user's claim of "no background tasks" might be inaccurate—hooks or MCP servers can trigger calls without explicit user action.

How to Fix It

1. Kill all Claude Code processes immediately:

   claude --kill-all
   

   This terminates every active session, including background ones.

2. Check your usage dashboard:
   Go to claude.ai/usage and look for any spikes or activity that doesn't match your actual usage. If you see usage when you're not working, you've found the leak.

3. Review your hooks and MCP servers:
   Check ~/.claude/settings.json for hooks that might run on events like SessionStart or PreToolUse. Disable any that aren't essential. Similarly, review your MCP servers—some can make autonomous calls.

4. File a support ticket with data:
   If the issue persists, contact Anthropic support with your usage logs and timestamps. The Reddit user found the support bot useless, but a detailed ticket with evidence can escalate to a human.

Why This Happens

Claude Code's architecture allows for long-running sessions that can survive terminal closure. If a session crashes or gets stuck in a loop, it may keep consuming quota. The 5-hour reset is a rolling window, so any background activity right after reset can instantly exhaust it.

Try It Now

• Run claude --kill-all before and after your work sessions.
• Use claude --resume to reattach to a session instead of letting it run detached.
• Monitor your usage daily at claude.ai/usage to catch anomalies early.

If you're on a Pro plan and hit limits frequently, consider switching to the API with a budget cap—you'll have more control over token spend.

Source: reddit.com

[Updated 06 Aug via hn_claude_code]

The issue may not be isolated to local background processes. A new report on Hacker News ties the rapid quota drain to a recent Claude status incident (status.claude.com/incidents/mgp99sn4ynd4), suggesting a possible server-side bug. The user, on a $100 Claude Code plan, also had the clau.de/cc-50-promo enabled, adding a promo-code variable. If Anthropic's incident caused misattributed usage, local fixes like claude --kill-all won't help. Affected users should check the status page and mention the incident ID when filing support tickets, as it may accelerate recognition of a broader issue. [per Hacker News]

[Updated 06 Aug via reddit_anthropic]

A separate Reddit report suggests the problem may be broader than background sessions. A Pro user on Opus 4.6 reports being limited to only two turns per session over the past 48 hours, with widening gaps between sessions, even for simple prompts. They question whether Anthropic has silently tightened account-specific limits for all Pro users. This aligns with the HN incident theory but adds a new symptom: not just instant quota exhaustion, but a visible reduction in per-session capacity. If limits are being adjusted server-side, local fixes like claude --kill-all or checking hooks won't address the root cause. [per Reddit]

[Updated 07 Aug via devto_claudecode]

A developer on the Max plan reports that 97.4% of their 1.36B weekly tokens were cache_read—not new usage. This suggests high token totals may be misleading, as cache reads cost about one-tenth of regular input. The user logged 13,972 turns in a week, averaging ~97,500 tokens per turn, yet only ~36M were new cache writes or output. This indicates that quota exhaustion may stem from excessive cache reads, not just background sessions. [per dev.to]