[BUG] Claude Opus 4 limit reached, now using Sonnet 4

[wysiecki]

Environment

• Platform (select one):
  • [ ] Anthropic API
  • [ ] AWS Bedrock
  • [ ] Google Vertex AI
  • [ X] Other:MAX 100$
• Claude CLI version: 1.0.16 (Claude Code)
• Operating System: MacOS 15.5
• Terminal: iTerm2

Bug Description

Getting on the first request, already the limit reached message, was not like this before, happening since two days. I am aware of limitations, but please, on first request in first session of the day ?

Steps to Reproduce

1. Not that complicated task 2.no MCP's used

Expected Behavior

Would love to use opus 4 a little bit longer on a 100$ subscription

Actual Behavior

Claude Opus 4 limit reached, now using Sonnet 4

[dicksontsai]

Can you share a screenshot (if possible) of when you saw the fallback and the output of /status?

While I'd say this is unexpected, there could be a variety of other factors based on input/output tokens, like what you include in CLAUDE.md, what files you read, what output you're having the model generate.

[hazmyr]

@dicksontsai @wysiecki

1. First task for today: ⏺ Task(Analyze security warnings) ⎿ Done (10 tool uses · 37.1k tokens · 1m 33.7s)
2. Second Task ⏺ Task(Check Ghost alignment) ⎿ Done (11 tool uses · 43.1k tokens · 3m 52.5s)
3. At third task for today limit reached:

Now let's update the todo list and run the tests:
  ⎿  ☒ Analyze security warnings and create detailed fix plan
     ☒ Create security utilities for validation and safe access
     ☒ Fix 20 object injection warnings in core files
     ☒ Remove 5 console statements from production code
     ☐ Run full test suite after security fixes

⏺ Bash(npm test)
  ⎿  Error: > [email protected] test
     > npm run lint:security && npm run test:ts && gscan .


     > [email protected] lint:security
     > eslint assets/js/**/*.js --no-ignore


     /Users/razvanmoraru/WorkPersonal/my-claude-project/content/themes/ghoststarter-master/ass
     ets/js/core/config-manager.js
       129:31  warning  Generic Object Injection Sink        security/detect-object-injection
     … +44 lines (ctrl+r to see all)

⏺ We have a few issues with the security.js file (formatting and new console statements). Let
   me fix those:

⏺ Update(assets/js/utils/security.js)
  ⎿  Updated assets/js/utils/security.js with 9 additions and 9 removals
     190              const noop = () => {};
     191 
     192              return {
     193                  log: isDebugMode() ? 
     194                      (...args) => console.log(`[${module}]`, ...args) : 
     195                      noop,
     196                  warn: isDebugMode() ? 
     197                      (...args) => console.warn(`[${module}]`, ...args) : 
     198                      noop,
     199                  error: isDebugMode() ? 
     200                      (...args) => console.error(`[${module}]`, ...args) : 
     201                      noop
     193                  log: isDebugMode()
     194                      ? (...args) => console.log(`[${module}]`, ...args)
     195                      : noop,
     196                  warn: isDebugMode()
     197                      ? (...args) => console.warn(`[${module}]`, ...args)
     198                      : noop,
     199                  error: isDebugMode()
     200                      ? (...args) => console.error(`[${module}]`, ...args)
     201                      : noop
     202              };
     203          },
     204 

Claude Opus 4 limit reached, now using Sonnet 4

claude.md has 192 line: structured like this

I hope this helps, didn't use today Claude app or browser with Opus, is strange since on the app it doesn't show the the limit was reached, also I don't think that this 3 commands or the tokes used are 20%, or if I'm wrong can you provide more details. Thank you!

[hazmyr]

I also uninstalled it and installed it again today, before starting this session

[dicksontsai]

Thanks for your detailed response @hazmyr . Based on what you have provided, I am not surprised that you would hit 20% of your Claude Max 5x usage limit with Opus 4. I do not see anything wrong with how the client is functioning.

Opus 4 is an expensive model, and it will use up your rate limits approximately 5x faster than Sonnet 4. That said, the performance can be well worth it. The https://www.anthropic.com/pricing#api page can give you an idea of the true cost of these queries relative to what you are paying with your current plan.

You can avoid the 20% fallback by running /model opus to use Opus all the time.

UI note: When you see

⏺ Task(Analyze security warnings)
⎿ Done (10 tool uses · 37.1k tokens · 1m 33.7s)

It is just reporting the number of tokens consumed by the Task (subagent) tool call. Your screenshots show other usage like the file writes for your planning and utility scripts, the tool calls to process the bash test output and subsequent file edits, etc.

[hazmyr]

@dicksontsai thank you for your quick reply! I'm a little bit confused about this "You can avoid the 20% fallback by running /model opus to use Opus all the time." Opus and Opus 4 are not the same?

[dicksontsai]

They are the same for now, but say there's a future Opus X, where X > 4. Then, Claude Code will use that newer Opus instead of staying pinned on Opus 4. /model opus (which we resolve for you) vs. /model claude-opus-4-20250514 (full model name, not recommended)

[hazmyr]

@dicksontsai got it, thank you for the quick response and the explanation! Solved on my side!

[dicksontsai]

Closing this issue because there are no other details that we can evaluate or take action on.

Our general advice is that:

1. You can pin the model to Opus (/model opus), and this defaults to Opus 4. Note that Opus reaches usage limits ~5x faster than Sonnet.
2. To maximize Opus time, the tips about "Managing Costs" are still relevant: https://docs.anthropic.com/en/docs/claude-code/costs

[github-actions[bot]]

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.