GPT 5.5 in Codex is actually crazy... I've had it churning for a few hours now on some pretty complex bio data pipelines, and it's proactively making code changes while it waits for data download commands to finish. I've never seen this behavior from Codex before this model