I've found Sol's propensity for delegating to subagents can make it... disastrously expensive, especially with each subagent having some implicit floor on further reasoning/context gathering before action.

The base model is certainly cheaper and more token efficient etc, but on large tasks cost in some way is now n^2

I have the codex x20 plan also. Set it to 5.6 sol ultra, gave it one task, in less than 10 mins I got the 10% warning. I did have it do a few things earlier, just some anaylsis and then create a pdf from it. I thought maybe that took a little longer than I though.

Used a reset, it went for about 11 minutes and then, just out of usage popped up, no warning, there were still 4 agents running, the nice thing is it did let them finish, each one went for about 10 minutes.

I also have a claude max plan, I have been using Fable 5 on ultra, I never hit the session limit, and get 3 or 4 full day's looping on ultra.

I don't know how it handles the subagents, but claude does it much more efficiently, Codex does seem much faster, so maybe it's just a relativity thing.

How bad is it for you? Are you on ultra or xhigh/max? I typically ask it (5.5, now 5.6-sol) to use subagents for specific things anyway. On the Pro 20x plan, I'm seeing like ~1% usage per 20-30 min per session (on max effort), which is in line with 5.5. Currently trying out ultra on a personal project, feels like ~3x more expensive per unit time. (No idea on quality yet, for obvious reasons.)

Max via Cursor, should've mentioned, very possible I'm seeing more Cursor than OAI. All the subagents it picked were also Sol Max... I've seen 7-10 in one turn. Regardless the top reasoning tiers being conflated with subagent count feels double edged

My global CLAUDE.md/AGENTS.md etc have strict rules to never use subagents without approval or when specifically requested except for whitelisted scenarios.

I think that's the way to go most of the time

imo there’s a clear greenfield to have doubled down on where cursor was before in proactively keeping devs appraised of the code that they’re generating, and bridging growing gaps between abstracted chat sessions and files/directory structures I might understand less and less

This on the other hand feels like a clear reaction to cc/codex, in a way that even kind of builds an offboarding ramp?

would sure be nice if the effort spent fighting their karma was pointed at a better frontier model

Many are saying codex is more interactive but ironically I think that very interactivity/determinism works best when using codex remotely as a cloud agent and in highly async cases. Conversely I find opus great locally, where I can ram messages into it to try to lever its autonomy best (and interrupt/clean up)

where we draw the line on agent "identity" when the models being orchestrated are generally the same 3 frontier intelligences is an interesting question indeed

I would think this idea of creating a third-party to verify things likely centers more around liability/safety cover for a steroidal increase in velocity (i.e. --dangerously-skip-permissions) rather than anything particularly pragmatic or technical (but still poised to capture a ton of value)

Appreciate the feedback.

Will extend the decision table template when we can

Guilty as charged

Thoughts on /permissions from the CLI?

The wording has been updated, and we’ll review our instrumentation. Appreciate your feedback.