Comment by elkoan
5 hours ago
I've yet to see Anthropic promote any sort of token optimization strategy to its users - they always assume we all have infinite inference.
"No bread? Let them eat cake!"
5 hours ago
I've yet to see Anthropic promote any sort of token optimization strategy to its users - they always assume we all have infinite inference.
"No bread? Let them eat cake!"
Not sure how you use CC, but the last 6 months has felt like significant optimization efforts to me. Last year Claude would just read and edit files, now it's all kinds of basic tool gymnastics with grep/awk/sed/etc to narrowly slice and avoid token-heavy reads. Resuming sessions that aren't even that large get a scary prompt about using a significant portion of your token budget if you continue without compacting.
To me it feels like a worse experience, and they probably feel it too, but it makes sense from an optimization perspective. I've probably learned some shell tricks, but also going blind from watching Claude try dozens of variations of some multi-line chained and piped wall of bash nightmare, instead of just reading a few files.
Valid points, but they address a totally different matter than the one I pointed out.
I completely agree but hadn’t found a way to put it to words. It could be the model too trained on optimized strategies
Nah they do. They push Sonnet pretty hard rather than Opus for most tasks.
Also: https://platform.claude.com/docs/en/agents-and-tools/tool-us...
I've noticed that's changed over the past month or so. Claude-code used to happily pipe build commands straight into context, but recently it's been running them as background tasks that pipe to file, and it'll search and do partial reads on the output instead.
It also gives tips on reducing context size when you run /context .
Presumably they are actually starting to feel the pinch on inference costs themselves with what still feels like a fairly generous max plan.
And it seems to use head, tail etc. more than it used to, even when unnecessary, which, combined with the recent(?) tendency of more chaining and as you said, piping to temp files and the like, totally screwed up claude code’s auto approval system for me (by auto approval I mean the system to decide which commands can be run without permission prompt, based on the permissions.allow setting among other things, not to be confused with a specific new approval mode called “auto” that burns more tokens to decide whether the command is safe). I had to write my own auto approval system and plug it in as a hook.