I've burned over $200,000 of tokens with gpt-5.6-sol. It's a great model. That said, it's a bit too easy to hit your limits on the $200 Codex Pro sub.
OpenAI has been generous about resets, but that doesn't help you when you kill your 5 hour window with 4 hours remaining. I made a lot of mistakes and see others making them too. I'll probably do a video on this eventually, but wanted to get this advice out ASAP so you guys can get more done while burning less.
Effort levels
Default to medium or high. Both are great. xhigh is incredibly capable but I don't find myself needing it much, even when orchestrating lots of subagents (more on this later)
“Ultra”
Ultra is not a reasoning level, despite where they put it in the UI. It's causing similar confusion to Claude Code's “Ultracode”. I have a video coming out about this later.
For now, I recommend avoiding Ultra entirely. There are bugs in the Codex harness that cause it to spawn way too many subagents with WAY too high of reasoning levels. I will revisit this when the bugs are resolved.
Fast mode
I love fast mode. I used to use it a lot. It makes sense for models that are prone to stopping a lot before burning a ton. Reminder that fast mode uses 2.5x as much credit.
5.5 would often stop and need encouragement to keep going. 5.6 can go for WAY LONGER. This is mostly a good thing. It can be trusted to compete tasks end to end. It also makes the “token burn” a lot less predictable.
Ignoring /goal, a single message with 5.5 could use between 0.1% and 2% of my limits. With the 2.5x, that was ~5% “peak” for a given message.
I've seen 5.6 use as much as 15% on a single message because it goes way way longer. With the fast mode multiplier, that would be 40% of your 5 hour window in a single message. I know this is burning a LOT of people. Trust me, don't use fast mode for a bit.
Subagents
This is the coolest unlock with GPT-5.6. It's also easy to footgun. Sol is very eager to spin up subagents. This is mostly good.
Sadly, the implementation in Codex has a whole rat's nest of issues (don't get me started on the v1/v2 split and auto-routing based on models).