claude-code - 💡(How to fix) Fix [Bug] Is Claude blowing through Max plans due to removing token output constraints? [1 comments, 2 participants]

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…
GitHub stats
anthropics/claude-code#46185Fetched 2026-04-11 06:26:53
View on GitHub
Comments
1
Participants
2
Timeline
6
Reactions
0
Timeline (top)
labeled ×4commented ×1renamed ×1

Root Cause

During this same period, I have seen Claude seriously producing way too much output compared to how it would provide replies in the past. Historically, Claude has been very constrained in its answers in Claude Code. That has absolutely changed in Claude Code. Set aside the token usage—I find interactive sessions to be much more mentally taxing because every answer feels 5x longer than it needs to be. That is a real change. One of the first things I remember striking me when I first used Claude Code was how much more constrained its prose was than in the web app.

RAW_BUFFERClick to expand / collapse

Bug Description

Many people have reported blowing through their Claude Code plans. It's definitely impacted me—and at a terrible time as I've been running a prompt injection benchmark suite in preparing an upcoming publication on mlld's effectiveness at prompt injection defense which chews through tons of tokens anyway.

During this same period, I have seen Claude seriously producing way too much output compared to how it would provide replies in the past. Historically, Claude has been very constrained in its answers in Claude Code. That has absolutely changed in Claude Code. Set aside the token usage—I find interactive sessions to be much more mentally taxing because every answer feels 5x longer than it needs to be. That is a real change. One of the first things I remember striking me when I first used Claude Code was how much more constrained its prose was than in the web app.

Claude thinks forever. I'm sitting waiting longer for it to think about most questions than I have seen it think when I used to use ultrathink!

And to be clear: I'm not even using high effort! I changed to 'auto' because high was producing massive amounts but that didn't reduce it.

<img width="321" height="52" alt="Image" src="https://github.com/user-attachments/assets/ef4828e9-4dc5-4e2e-8c29-afe6527371cf" />

Rather than speak anecdotally, I have two recent cases where I've captured a section of output in gists.

1. Claude double answers

In this example, you can see Claude answer the same question twice at significant length to what was a trivially easy-to-answer question. I have this double-answer thing happen frequently.

2. Claude loses it

In this example, Claude goes off the rails (it's not even responding with anything close to what the topic of the conversation was) and is also unable to stop.

It's pages of this nonsense, but something Claude said in the middle of that nonsense actually made me realize that is what I'm seeing: Claude is just over-outputting everything lately.

<img width="912" height="729" alt="Image" src="https://github.com/user-attachments/assets/ab9dd5d5-8fd8-42fb-ba61-50f3968ef4fe" />

Shared here for visibility and entertainment purposes:

https://x.com/sockdrawermoney/status/2042550449417654759?s=20

Environment Info

  • Platform: darwin
  • Terminal: ghostty
  • Version: 2.1.98
  • Feedback ID: 0bce4945-99ad-4e9a-b85c-fa22a8e52297

extent analysis

TL;DR

Adjusting the effort level or exploring alternative configuration options may help mitigate the issue of excessive output from Claude.

Guidance

  • Review the provided gists to understand the patterns of excessive output and double answering, which may help identify specific triggers or scenarios that contribute to the issue.
  • Experiment with adjusting the effort level, as the user has already tried switching from 'high' to 'auto' without significant improvement, to find a balance between response quality and output length.
  • Investigate if there are any recent changes in the Claude Code environment, such as updates or modifications to the underlying models, that could be contributing to the increased output.
  • Consider reaching out to the developers or support team for Claude Code to report the issue and gather more information about potential fixes or workarounds.

Notes

The issue seems to be related to changes in the behavior of Claude Code, particularly in terms of output length and response quality. However, without more information about the underlying models or recent updates, it is difficult to provide a more specific solution.

Recommendation

Apply workaround: Adjust the effort level or explore alternative configuration options to mitigate the issue, as the root cause of the problem is unclear and may require further investigation or support from the developers.

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING