I mean, not exactly first but still

I’m now on ChatGPT Plus! I’ve already used Codex before, on the free plan, but this’ll be the first time using it on a paid plan. I’ve immediately put it to work fixing a few things Claude left in-flight, and it did them without much complaint on Terra.

One thing I’ve noticed is that the remote connection functions a bit differently - on Claude Code I could use either the CLI (/rc) or the app. Codex wants me to use the ChatGPT app (installed with paru - I seem to be liking the AUR a lot). It also seems to maintain a connection even when my phone isn’t viewing the mobile app, as it shows a persistent notification saying it’s connected. Reminds me of Termux staying connected, not sure how it’ll impact battery use! Finally, threads where Codex is running in the CLI I don’t seem to be able to “attach to” via the remote connection. Would be nice if it at least errors out cleanly, because for me it seems to kind of stall.

The remote control generally seems both more polished and less. It took a while to reconnect after I switched to mobile data, but also it seems to stream results and in between steps better - might be a trade-off of whatever method they’re using for this connection.

I’ve also noticed the 256k-ish default context limits. It is apparently both expandable to 1M and not recommended, and it’s applied on both Terra and Sol - I’d imagined they wouldn’t have issues with having Sol (their previous flagship) use the full available context, but apparently they’d prefer even it stayed under that limit. Possibly related to their API pricing, which, if it applies to Codex quotas too, would mean it’ll burn through my usage limits extra quickly!

On rate limits, usage limits and other more detailed factors… I’ll have to see. Terra seems very generous, Sol less so, I’m scared to trigger Astra except as a last resort / orchestrator only. OpenAI apparently allows Codex agents to keep working up to abuse prevention limits fair-use limits when I hit a quota limit (Corrected by an LLM proofreading pass — OpenAI’s documentation says the active turn may continue subject to fair-use limits, rather than “abuse prevention limits.”). I’ll continue using it, and will update you as I find out new things! I still have to wonder if anyone is watching now…