Fine but please don't let this become another thread where everyone gets told to tune their workflow. Colleague of mine hit the cap twice at 16% usage on SHORT sessions. Screenshots, timestamps, the lot. If your numbers don't add up, that's not a you problem and the tuning advice is noise.
Why does my Claude Max usage run out so fast in Claude Code?
Posted by rachelops · Jul 22, 2026
Two different things get mashed together in every version of this thread and it makes the whole topic useless to search.
Thing one: people hitting the limit at like 15% of stated usage, minutes in, with screenshots. That's a real reported defect, there's an open issue, go add your numbers to it. Nothing below will help you.
Thing two, which is most of us: long session + big context + lots of tool calls is just expensive, and it's expensive in a way that isn't obvious, because you're not paying for what you type. You're paying for what gets re-sent. Every turn ships the whole conversation again. Carry 400KB of context for three hours and you paid for that 400KB on every turn in those three hours.
Stuff that actually moved our numbers:
/clearbetween unrelated tickets./compactis not the same thing, it summarises and carries on, the floor stays high.- stop pasting files in. let it read them. pasted content lives in the transcript forever, tool-read content you can clear.
- the killer pattern is: agent greps around, reads 40 files, finds the thing in file 3, then carries all 40 for the rest of the day.
- lower effort for mechanical stuff.
Didn't help: shortening prompts (prompt is the cheap part), switching model mid-session (worse, blows the cache).
Anyway I'm curious what everyone else's shape looks like. One long session all day, or lots of short ones?
Discussion · 7 comments
Strong evidence gets promoted into the record above.
The pasting thing. I had a 3000 line schema I pasted into message 1 of every session, out of pure habit, for months. Every turn after that re-sends it. Nobody tells you this anywhere.
edit: to be clear I mean it was in my first message, so it's in the transcript, so it's in every subsequent request. reading it with a tool instead was the single biggest change I made.
many short ones here, ~6-8 a day, and I basically never hit the limit. so anecdotally this tracks.
Sign in to join the discussion, vote, and verify fixes.
Related pages
one long session, all day, and I think that was my entire problem. I kept it running because restarting felt wasteful, all that context I'd built up. had it completely backwards. the context I thought I was saving was the thing billing me. clearing between tickets roughly halved my usage and I genuinely cannot tell a difference in output.
same shape here. the "warmed up session" thing is a trap, it is warmed up and you're paying rent on it every single turn.
does /clear lose the CLAUDE.md too or does that get reloaded
reloaded. that part's fine.