• AgentArmor teamTeam3d agoNeeds helpSolved

    Claude Code 5-hour rate limit exhausted in about 70 minutes after v2.1.89, does v2.1.90 fix it?

    On a Max 20 plan using Claude Code 2.1.89 on Ubuntu 24.04 with Opus 4.6, the 5-hour usage limit reached 100% within about 70 minutes of light work after a reset, where earlier versions lasted 3-4 hours. The reporter asked whether this was a client-side regression and how to avoid it.

    Originally asked on GitHub: https://github.com/anthropics/claude-code/issues/41788

    6 answers

    • Best answerAgentArmor teamTeam3d ago

      The status is fixed on the client side according to commenter 12, who says v2.1.90 largely fixed the cache regression behind their drain. Server-side quota concerns raised in the thread were not resolved.

      Steps the thread reports:

      1. Update to v2.1.90 with claude update.
      2. To stop auto-updates from changing your version, add this to the env block in ~/.claude/settings.json:
      json
      {
        "env": {
          "DISABLE_AUTOUPDATER": "1"
        }
      }
      
      1. Commenter 12 reported that the npm install (npm install -g @anthropic-ai/claude-code) gave slightly better sub-agent cache performance than the standalone binary on v2.1.90.

      Commenter 10 used a monitoring proxy set through ANTHROPIC_BASE_URL on v2.1.89 and measured cache read ratios as low as 4.3% across 168 turns. Their workarounds raised the ratio to 89-99.5%. These are their measurements, not confirmed by Anthropic.

      Commenters 3 and 4 reported that downgrading to v2.1.88 or v2.1.68 helped, but they did not provide measurements.

      General advice: avoid --resume on long sessions and run fewer parallel sessions, since commenter 11 links these to extra token use.

      Did this work for you?
    • AgentArmor teamTeam3d ago

      What a GitHub user tried on GitHub (Didn't work)

      Reporter on Claude Code 2.1.89 trimmed CLAUDE.md and MEMORY.md from about 30KB to 14KB and checked for background processes and concurrent sessions. The rate limit still drained quickly.

      Source: https://github.com/anthropics/claude-code/issues/41788 (opens external site)

      Did this work for you?
    • AgentArmor teamTeam3d ago

      What @karytabilu tried on GitHub (Outcome unclear)

      Commenter downgraded to Claude Code 2.1.68 and reported that token drain stopped.

      Source: https://github.com/anthropics/claude-code/issues/41788#issuecomment-4167694644 (opens external site)

      Did this work for you?
    • AgentArmor teamTeam3d ago

      What @minodevss tried on GitHub (Outcome unclear)

      Commenter downgraded to Claude Code 2.1.88 and believed the problem was solved.

      Source: https://github.com/anthropics/claude-code/issues/41788#issuecomment-4167700606 (opens external site)

      Did this work for you?
    • AgentArmor teamTeam3d ago

      What @ArkNill tried on GitHub (Worked)

      Commenter ran a monitoring proxy via ANTHROPIC_BASE_URL on 2.1.89, measured a cache read ratio as low as 4.3%, and applied cache workarounds that brought the ratio to 89-99.5%.

      Source: https://github.com/anthropics/claude-code/issues/41788#issuecomment-4168818478 (opens external site)

      Did this work for you?
    • AgentArmor teamTeam3d ago

      What @ArkNill tried on GitHub (Worked)

      Commenter benchmarked v2.1.90 standalone and npm builds and reported cache reads of 95-99%+ with much lower usage. Recommended running claude update and setting DISABLE_AUTOUPDATER to 1 in ~/.claude/settings.json env.

      Source: https://github.com/anthropics/claude-code/issues/41788#issuecomment-4174665270 (opens external site)

      Did this work for you?

    Add your answer or comment