Tony ArnoldJan 8

I switched over to using Claude Code this morning to do some work, and it promptly used 50% my usage for the next 5 hours resolving a simple build failure. I have my model set to Sonnet 4.5 and it's a Pro plan, so it's not meant to burn usage that quickly…

There's something wrong with it, right? Codex just happily burbled away fixing shit like this all of yesterday and never once got close to the usage limits.

Show threadJamesJan 8
@tonyarnold I’ve been using Claude Code to play around with some stuff using that super power skill. It’s pretty common that I can’t get it to build one (smallish) screen without blowing through the usage and having to finish it later. Interesting to know Codex doesn’t have that problem
Show threadTony ArnoldJan 8

@JamesTech I used the superpowers tool with Codex most of the past week (since I posted about it) — slightly higher usage, but it worked great.

Switched to Claude Code and it's ridiculous — I am beginning to look at whether it's interacting with something else I have installed like Axiom.

Show threadTony ArnoldJan 8
@JamesTech yeah, it was (partially) Axiom. Uninstalling that plugin makes everything functional again, but it is still burning through usage at a rate of knots (75% in ~25 minutes).
Show threadJamesJan 8
@tonyarnold Interesting, that pretty much matches what I’ve seen too with usage
Show threadTony ArnoldJan 8

@JamesTech weird. I get hours out of Codex with GPT-5.2-Codex set to high reasoning.

Maybe this is just how Claude Code is?

Show threadJamesJan 8
@tonyarnold Yeah maybe. Very possible I’m just holding it wrong but 30-40 minutes is about all I get with a fairly simple setup. Dash for documentation, XcodebuildMCP, and that superpower plugin. This is on a toy project too with less than a few thousand lines of code. Might have to give Codex a go
Show threadChris KolbuJan 8
@JamesTech @tonyarnold it makes sense that «usage limits» are used as an opaque lever to pull to adjust inference spend.
Show threadTony ArnoldJan 8

@nesevis @JamesTech the whole thing feels like a rug pull waiting to happen, obviously.

There's no meaningful correlation that's been publicly communicated about how, when or why the usage limits are determined.

I look forward to a point where:

1. Usage limits are far higher for similar cost to the end user; or
2. We can run some variation of LMs meaningfully, locally.

Show threadChris KolbuJan 9
@tonyarnold @JamesTech I’d love to be able to run something useful locally. It would then also become a meaningful second life for older (still-capable) machines.
Show threadLuke Sleeman

@nesevis I'm also watching the "run things locally" space closely. It seems like the only way to really avoid the mess that's going to happen when the bubble bursts and the surviving companies stop selling usage at a loss.

I don't think things are quite there yet, but it does look promising. Particularly if a bunch of companies go belly up, and suddenly the market is flooded with unused GPUs

Jan 9 at 12:44amWeb