← sources

"Opus usage is too high" - Advise for you

source: reddit (reddit:ClaudeAI)published 2026-10-04 16:08ingested 2026-10-05 02:37

retained text

Hi there,

I've seen several posts with different experiences when it comes to usage and intelligence. My experience is good in general but today I've found that 25% of my weekly usage of my x20 subscription was gone. I left Claude working during the night and I wasn't at home during the morning.

So did Opus usage increased? No.

What to do now? I asked Claude what was going on specifically with agents using a lot of tokens and staying up for so long.

Answer:

    What is really wrong
    - The agents wait a lot. Tests from 5 agents share one line, so an agent can wait a long time before its test even starts.
    - The agents use too many tokens. Each one used 750k to 945k tokens. That breaks your rule to use tokens well. The cause is agents that live too long: every step re-reads all of their old work.
    What I recommend
    - When each agent says "CHECKPOINT OK", I stop it.
    - After your restart, I start new, small agents. Each one does one job from the handoff note (its "what is done, what is next" file), on Sonnet where the job is easy.
    - Each agent runs only the 1 or 2 tests for its own job.
    That fixes the token waste and the waiting. Is that OK?

Conclusion

This was probably because Claude ignored some of my rules on [Claude.md](http://Claude.md) about spawning Sonnet agents for easier tasks and how to handle certain tests. The tokens are gone, it sucks but at least I know why and it's solved now. I just want to share my experience so if it happens to you, you can also do something about it.

Take care everyone

document fetched 2026-10-05 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1wxjdxm/opus_usage_is_too_high_advise_for_you