[BUG] API Error: Claude's response exceeded the 32000 output token maximum.

[patrikhuber]

Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report (please file separate reports for different bugs)
• I am using the latest version of Claude Code

What's Wrong?

Hi,

I just asked Claude Code to implement something relatively hard. After 8 minutes of thinking, there is this error in the Claude Code terminal window:

⎿  API Error: Claude's response exceeded the 32000 output token maximum. To configure this behavior, set the
CLAUDE_CODE_MAX_OUTPUT_TOKENS environment variable.

However it still seems to be thinking (14 minutes now), so it's probably still going. It hasn't given me any feedback of what it's doing or thinking.

✽ Shenaniganing… (11m 35s · ↓ 35.0k tokens · thinking)

Claude Code on PowerShell on Windows Terminal, Windows 11, latest version (updated today). Using Opus 4.6. On the Pro plan currently.

I found a few related issues (other users encountering this issue) but they are all closed or in a different context.

What Should Happen?

No error, or a better explanation.

Error Messages/Logs

API Error: Claude's response exceeded the 32000 output token maximum. To configure this behavior, set the
     CLAUDE_CODE_MAX_OUTPUT_TOKENS environment variable.

Steps to Reproduce

I can share my prompt privately if you want to contact me. But I suspect this is reproducible with any prompt that requires long thinking (very complex task).

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

No response

Claude Code Version

2.1.37 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Windows Terminal

Additional Information

No response

[carrotRakko]

Same error here on Opus 4.6 (v2.1.37). I also tried setting CLAUDE_CODE_MAX_OUTPUT_TOKENS=64000 but the 32000 limit remained unchanged. Filed a separate report for that: #24159

✍️ Author: Claude Code with @carrotRakko (AI-written, human-approved)

[patrikhuber]

I also find the documentation (https://code.claude.com/docs/en/settings) confusing:

CLAUDE_CODE_MAX_OUTPUT_TOKENS | Set the maximum number of output tokens for most requests. Default: 32,000. Maximum: 64,000. Increasing this value reduces the effective context window available before auto-compaction triggers.

First, this docs page (https://platform.claude.com/docs/en/about-claude/models/overview#model-comparison-table) says Opus 4.6's output limit is actually 128K, so above the max of 64k that the other docs page lists (perhaps this just hasn't been updated for the Opus 4.6 release).

Second, I'm generally confused as a user what to do. It sounds like to make this error go away, I have to set CLAUDE_CODE_MAX_OUTPUT_TOKENS to 128K. But then increasing this value reduces my context window - so I probably don't want to do that? (I also don't understand why increasing the max output would reduce my context window?).

Would be nice to make this clearer/easier for the user.

[nullbio]

This is driving me insane. I'm using Codex in the mean time. CC literally unusable right now. It takes 15 minutes to do a single task at the best of times, and we have no visibility into what is happening. I used a packet inspection tool to see if I could figure out where it's getting stuck, and quite often it's just sitting on an open TCP connection waiting for SSE events to come back from Anthropic. In other words, we're just waiting in a queue, forever.

In a more related case, it was sitting there for 12 minutes and 30 seconds to redesign a landing page, all it did was adjust some wording and colors. There's major problems and no one seems to be speaking about it. Genuinely thinking about cancelling my max sub at this point because I can't use CC like this and Anthropic seems to just ignore everyone so I doubt it'll get fixed any time soon.

● Direction confirmed.

  Let me build it.
  ⎿  API Error: Claude's response exceeded the 32000 output token maximum. To configure this behavior, set the CLAUDE_CODE_MAX_OUTPUT_TOKENS environment variable.

● Writing the redesigned page now.

12 minutes to get to that API error after "Direction confirmed".

[stefansimik]

I also find the documentation (https://code.claude.com/docs/en/settings) confusing:

CLAUDE_CODE_MAX_OUTPUT_TOKENS | Set the maximum number of output tokens for most requests. Default: 32,000. Maximum: 64,000. Increasing this value reduces the effective context window available before auto-compaction triggers.

First, this docs page (https://platform.claude.com/docs/en/about-claude/models/overview#model-comparison-table) says Opus 4.6's output limit is actually 128K, so above the max of 64k that the other docs page lists (perhaps this just hasn't been updated for the Opus 4.6 release).

Second, I'm generally confused as a user what to do. It sounds like to make this error go away, I have to set CLAUDE_CODE_MAX_OUTPUT_TOKENS to 128K. But then increasing this value reduces my context window - so I probably don't want to do that? (I also don't understand why increasing the max output would reduce my context window?).

Would be nice to make this clearer/easier for the user.

I have exactly the same question

[alalalalalalalaal]

Same issue, even changing the variable value seems to do nothing

[stefansimik]

Is it possible, that the variable CLAUDE_CODE_MAX_OUTPUT_TOKENS is applied when used API, but not when using subscription?

[old8man]

Yeah, it's a major issue. First we had a 128K limit (Max plan), then it shrank to 64K, and now we're slumming it at a measly 32K. I’d get it if claude code or Opus were actually designed to write to new files in bite-sized chunks. But right now, the model is just mindlessly incinerating tokens, utterly failing to save a single large file of its own creation. And to top it off, it repeats this little circus act two or three times, cheerfully flushing my paid subscription tokens down the drain. It’s absolute, unadulterated nonsense. Anthropic, have a heart )

[srb3]

Same issue for me, I tried changing the env var value but it did not seem to fix the issue