Sub-agent costs are too high and too opaque

Resolved 💬 5 comments Opened Feb 22, 2026 by MrLightspeed Closed Mar 29, 2026
💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

WTF!

I’m seeing unexpectedly high costs whenever sub-agents are used in Codex CLI. The feature is useful, but the current cost behavior makes it hard to use safely in normal development workflows.

Main issues:

  • Sub-agent usage can multiply spend very quickly.
  • Cost impact is not clearly shown before launching sub-agents.
  • It is hard to predict final cost from the prompt or plan.
  • There are not enough guardrails to prevent accidental overspend.
  • I am easily $350 over my Pro Plan for a week.

## What I need:

  • A global session budget that hard-stops execution well before you all tell me, "Sorry we have to screw up your job and tell you to buy credits" _on Day 2 of the week for a Pro User_.
  • A post-run cost breakdown by sub-agent , because I have no way to manage that unless _I totally turn off the feature._

Bottom Line

  • You are showing off sub agents while draining pro users, and pretty much everyone, while claiming they are cheaper. Yes, they really are cheaper, but you put the onus of managing that on us without cost controls.

I want to keep using sub-agents, but right now the pricing transparency and controls are not sufficient for practical day-to-day use. Please prioritize cost visibility and budget controls.

View original on GitHub ↗

5 Comments

github-actions[bot] contributor · 6 months ago

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

  • #12487

Powered by Codex Action

MrLightspeed · 6 months ago

NO

6639835 · 6 months ago

Hey, how did you count your costs? Afaik OAI uses credits instead of dollars

etraut-openai contributor · 6 months ago

Subagent token usage are not charged differently from other (main agent) tokens. Subagents allow you to do more in parallel, which will increase utilization. If you use the same model for subagents, the increase should scale linearly with the number of subagents you create. Are you observing something different? Depending on your use case, you may be able to use a smaller (and therefore cheaper) model for your subagents.

Can you clarify what you're asking for in this feature request? You mentioned a "global session budget". Can you explain more about how that would work?

etraut-openai contributor · 5 months ago

Closing for lack of response. There is some overlap with #14642.