Offtopic

01 April 2026 00:22

So not exactly 16 times slower, but still close.

Offtopic

01 April 2026 00:21

4 performance cores with 2 threads each, 8 ecores with one thread. 16 threads.

Offtopic

31 March 2026 23:52

I have a lot of cores.

Offtopic

31 March 2026 23:51

Anyway, I do not have 1 core.

Offtopic

31 March 2026 19:13

I'll only investigate if it happens again.

Offtopic

31 March 2026 17:45

It can compile packages, that would have made sense, but it didn't get that far.

Offtopic

31 March 2026 17:40

htop was sorted by CPU usage, so I couldn't see. Just that nix, the process I started before it began, was consuming about 20G.

Offtopic

31 March 2026 17:34

All cores were fully busy, but that was a side effect of the memory being full. As if was doing something special. But somehow I managed to get it full.

Offtopic

31 March 2026 17:31

100%, there is something called Caching. When you first talk to the LLM, it needs to know everything about your project.

The context window size doesn't change when you change the model, but most of the context for the first LLM is already cached on the server. When the next LLM is called, the whole thing has to be sent again, and the LLM has to know the proper context.

This is why you see that the amount of tokens used for the newer model is much higher.

Offtopic

31 March 2026 17:16

Maybe fair amount of token gets burned in the beginning

Offtopic

31 March 2026 17:15

Or to say it politely, doesn’t work for me.

Offtopic

31 March 2026 17:13

They should have compared Claude Code CLI with Claude Code in Cursor.

Offtopic

31 March 2026 17:09

I see. And that makes sense.

Offtopic

31 March 2026 17:08

No, not really... if you switch from one model to another, you have to send that new model all the context, while if you stick with a single model, it uses loads of cache.

Offtopic

31 March 2026 17:07

I mean there's an overhead I'd assume, but it's not 5 times as much.

Offtopic

01 April 2026 00:22

Which to the os looks like 16 cores.

Offtopic

31 March 2026 23:53

And they all work together. The fact my cpu may sucks is another argument.

Offtopic

31 March 2026 23:51

Matter of facts, I have 4 pcore and 8 ecore.

Offtopic

31 March 2026 22:58

👋

Offtopic

31 March 2026 18:19

repeating activities that lead to utilisation might have some insight. But it could be spontaneous 🤔

Offtopic

31 March 2026 17:41

For a package manager, certainly.

Offtopic

31 March 2026 17:39

That’s how we are. Makes sense totally

Offtopic

31 March 2026 17:33

My computer crashed today. I was lucky enough to have been able to open htop before it stopped reacting completely.

Offtopic

31 March 2026 17:16

To understand repo. It’s like complaints of diminishing returns is being improved

Offtopic

31 March 2026 17:15

And cause it’s due to caching.

Offtopic

31 March 2026 17:14

I’ve seen significant less token usage by Claude cli. Extension is useless.

Offtopic

31 March 2026 17:11

Comparing Opus with GPT-5 isn't a fair comparison.

Offtopic

31 March 2026 17:08

Token efficiency matters for cost. Independent testing found Claude Code uses 5.5x fewer tokens than Cursor for identical tasks. Claude Code (Opus) completed a benchmark task with 33K tokens and no errors. The Cursor agent (GPT-5) used 188K tokens and hit errors along the way. Fewer tokens means lower per-task cost even at higher subscription prices.

— https://www.builder.io/blog/cursor-vs-claude-code

Offtopic

31 March 2026 17:08

Claude models are good at coding so they gatekeep. No wonder. Their claims

Offtopic

31 March 2026 17:07

It does not seem to be the case for vscode.