They really need to figure out the rules, look I'd love to use a custom harness with Claude Code that I can extend, or build my own (which I'm doing) and use it with my Claude Code license, I don't want to overspend on tokens if I can help it. They really need to set their bar for the next model releases to use less tokens, or to trim their own cost for how these models are run. I'd be okay with a slightly slower experience with Claude Code if it meant similar throughput, but less cost, especially if I can build my own harness for it.

You can't really 'figure out' the rule that imposes restrictions on a client side interface of your models. As long as cli exists, third party software will always be able to interact with it, whether Anthropic likes it or not.

This is why it's so inconsistent and confusing. They simply can't come up with a rule that only affects OpenClaw/pi/etc and not 'allowed' automations. You either permit automation, or you don't. Right now, they want to have it both ways.

They might've just hit a ceiling with the quality they can get per token? Maybe the only real way left to scale quality is to increase token usage?

Out of curiosity, what kind of custom harness are you thinking?

> They really need to figure out the rules

A company built groundup on rule-breaking? Ain't gonna happen.