"We codex maxxed" twice in September. What's actually burning your credits?
We're a heavy user of Claude Code and Codex, and September got away from us. We moved Codex to the enterprise plan mid-month, burned 256K credits in the first two weeks, topped up 150K more, and ten days later half the team hit the wall the same afternoon. Slack that day was "we codex maxxed," "can't even use my reset," and "I need a bit of juice to continue a session, more please." One engineer gave Claude his .codex folder so it could reconstruct what he'd been doing and he could keep going on the other tool. Claude side, same week: "I'm going to hit my quota in the next 24 hours, can I get a bump."
When we finally pulled the numbers: one session had 3.2B input tokens across 2,400 writes. Sessions that touched gpt-6-astra averaged about $21. Same kind of work on gpt-5.6-terra averaged about $1. Astra was 38 percent of the Codex bill in 31 person-days.
Nobody did anything wrong. That is just what the defaults do when nobody can see them.
So, questions for others who live in these tools all day:
What was your most expensive session, and what was it actually doing? Context bloat, retries, a model that didn't need to be the big one, something dumber?
And what bugs you more day to day: hitting limits mid-task, a bill you can't explain, or not knowing which model to use for what?
Quick disclosure, I work at a startup that builds AI tooling. Happy to say which if anyone asks, but not pitching anything.