Started looking into this after I realized I was hitting usage limits faster than normal.

I find these subscriptions so essential to using LLMs for personal projects -- it would be prohibitively expensive otherwise

Codex, especially with Luna is an excellent deal. I think it's significantly more economical than Claude Code. The only reason to stay with CC at this point is access to Fable for _very_ hard problems.

Pulling the rug very slowly so that no-one notices and finding new ways to blow up your usage limits very quickly:

First it was "search agents" consuming the limits.

Secondly, then it was "reasoning tokens" doing that.

Third, it was "planning" mode for long horizon tasks.

Forth, adding "/goal mode" + 1M context for auto non-stop token consumption.

Fifth, "sub-agents" in parallel doing all of the above.

Then, Anthropic told us to use "loops" instead of prompting.

Recently "auto mode" was switched on by default, with sub-agents used by default on Fable.

Now, Anthropic is slowly reducing your limits in their closed-source harness.

Don't really understand what the drama here is. They said it was a temporary uplift and even though they could just take the whole uplift away, they just cut it by 25% which is better than before.

I really think Anthropic deserves much shit, but that's just shitting on the for the wrong reasons. And I hate that I am defending them here.