Unexpectedly high Codex Ultra usage in VS Code — is cost calculated relative to the user’s remaining allowance?

Over the past few days, I have been using the Codex extension in VS Code quite intensively, and I have noticed that Codex Ultra consumes an unusually large amount of my usage allowance.

When I was on the $100/month plan, a single moderately complex coding task could consume around 30% of my allowance. I later upgraded to the $200/month plan, which appears to provide roughly four times the usage allowance, but a similar task can still consume around 30–40%.

This makes me wonder: is the displayed usage percentage calculated relative to each user’s current balance or plan, rather than reflecting a fixed amount of compute or tokens? Otherwise, I would have expected the percentage consumed per comparable task to decrease significantly after upgrading.

I have also noticed several behavioural issues with Codex Ultra that seem to make the consumption much worse:

  • It repeatedly audits or rechecks code that it has already reviewed.
  • After completing the requested change, it often continues making unnecessary architectural changes.
  • It sometimes expands a narrowly scoped task into a much larger refactor, even when explicitly instructed not to do so.
  • Despite spending a lot of time on these extra steps, it still does not always follow the original instructions completely.
  • It can repeatedly inspect, modify, and re-audit the same areas, causing the allowance to be consumed extremely quickly.

The problem is not simply that Ultra uses more compute. The issue is that a substantial part of the usage appears to come from repetitive or unnecessary work that I did not request.

Has anyone else experienced similar behaviour? Are there any effective prompts, VS Code settings, workflow patterns, or task-scoping methods that can prevent Codex Ultra from repeatedly auditing the repository or making unnecessary architectural changes?

I would also appreciate clarification on how the usage percentage is calculated across different subscription plans.

Ultra is “multi agent” mode and will consume a lot of tokens. I would drop down to High on Sol to get to a more equivalent behaviour to 5.5 (although imho Sol is more tenacious and will keep going for longer)

Ultra on Terra is also an option (due to reduced consumption) but probably also not peak efficiency for most tasks.