AgentArmorHQ
Sign in
  • AgentArmorHQ teamTeam4d agoNeeds helpSolved

    Claude Code usage limits draining fast after March 23, 2026: which version and workarounds help?

    Since late March 2026, Claude Code users on Pro and Max plans report single prompts using a large share of their session quota and five-hour windows running out in under an hour. Users asked whether the cause is client-side and which version or settings reduce the drain.

    Originally asked on GitHub: https://github.com/anthropics/claude-code/issues/41930

    5 answers

    • Best answerAgentArmorHQ teamTeam4d ago

      Status: partly addressed, not fully resolved. Per the thread, the prompt-cache bugs were fixed client-side in v2.1.90 and v2.1.91, but commenters report server-side and accounting factors that remain unresolved.

      Steps that commenters reported helped:

      1. Update Claude Code: run claude update, or npm install -g @anthropic-ai/claude-code. A commenter reported 95-99% cache reads in stable sessions on v2.1.90.
      2. If you need to stay on a version, pin it by adding this to the env section of ~/.claude/settings.json: "DISABLE_AUTOUPDATER": "1".
      3. Avoid --resume. Commenters say it replays the full conversation as billable input and remains broken. Start fresh sessions instead.
      4. Limit parallel terminals and sub-agents, since each one draws from the same quota.
      5. Keep CLAUDE.md small, because it is resent as input on every turn.

      One commenter reported that v2.1.89 still had the cache bug and that v2.1.81 with a cache patch worked for them. Another reported that the 200K tool result cap may explain rapid drain in heavy coding sessions on v2.1.91. These are unverified reports, not confirmed fixes.

      General advice: if drain persists after updating, file a support ticket with your version and usage details. Server-side limit and billing questions are not fixed by client changes.

      Did this work for you?
    • AgentArmorHQ teamTeam4d ago

      What @ArkNill tried on GitHub (Partly worked)

      Reporter on a Max $200 plan identified a cache sentinel bug in the standalone binary and a resume bug from deferred_tools_delta since v2.1.69, and linked a community client-side workaround repo.

      Source: https://github.com/anthropics/claude-code/issues/41930#issuecomment-4169158245 (opens external site)

      Did this work for you?
    • AgentArmorHQ teamTeam4d ago

      What @ArkNill tried on GitHub (Partly worked)

      Commenter pinned v2.1.81 with a cache patch, used fresh sessions instead of --resume, used local proxy monitoring via ANTHROPIC_BASE_URL, and kept to one terminal with minimal sub-agents.

      Source: https://github.com/anthropics/claude-code/issues/41930#issuecomment-4173686349 (opens external site)

      Did this work for you?
    • AgentArmorHQ teamTeam4d ago

      What @ArkNill tried on GitHub (Partly worked)

      Commenter recommended running claude update (or npm install -g @anthropic-ai/claude-code), setting DISABLE_AUTOUPDATER to 1 in ~/.claude/settings.json env, and avoiding --resume. Reported 95-99% cache reads on v2.1.90.

      Source: https://github.com/anthropics/claude-code/issues/41930#issuecomment-4174668294 (opens external site)

      Did this work for you?
    • AgentArmorHQ teamTeam4d ago

      What @ArkNill tried on GitHub (Outcome unclear)

      Commenter tested v2.1.91 with proxy-based analysis, reported the cache layer fixed in v2.1.90-91, and suggested the 200K tool result cap may explain rapid drain in heavy coding sessions.

      Source: https://github.com/anthropics/claude-code/issues/41930#issuecomment-4181330348 (opens external site)

      Did this work for you?

    Add your answer or comment