[elephanlemon]

Yesterday (pro plan) I ran one small conversation in which Claude did one set of three web searches, a very small conversation with no web search, and I added a single prompt to an existing long conversation. I was shocked to see after the last prompt that I had somehow hit my limit until 5:00pm. This account is not connected to an IDE or Code, super confusing.

[master_crab]

Tool calls (particularly fetching for context) eats the context window heavily. I explicitly send MCP calls to sub agents because they are so “wordy”.

[bensyverson]

Everyone who has not hit this bug thinks it’s user error… It’s not. It happened to me a few days ago, and the speed at which I tore through my 5 hour usage cap was easily 10x faster than normal.

Also: sub agents do not get you free usage. They just protect your main context window.

[dmd]

I'm on Max. This morning, just to test, before doing anything else whatsoever, I was at 0%, and I typed 'test one two three' into CC.

That put me at 12%.

I have no MCPs except the built in claude-in-chrome.

This is clearly a bug.

[cyanydeez]

Readimg through this thread, it seems likely is a KV cache "bug". Theyre likely doing too many evictions of the LLM cache so the context is being reloaded to often.

Its a "bug" because its probably an intended effect of capturing the costs of compute but surfacing a fact that they oversold compute to a situations where they cant keep the KV cache hot and now its thrashing.

[bensyverson]

Caching helps them too, so I hope they fix it

[piva00]

Don't they consume less of the token quota in case the subagents are running cheaper models like Sonnet and Haiku compared to Opus?

[bensyverson]

Correct—I just wouldn't want folks to mistakenly think that the context fill % corresponds 1:1 with session token use.

[master_crab]

Yes, sorry. I meant it more as a descriptor of how many tokens it consumes. You are still stuck burning money.