›_ Codex Issues
Codex Issues / #23068

Token usage draining ~10% per prompt since May 10, 2026

Resolved 💬 5 comments Opened May 16, 2026 by sanrishi Closed May 16, 2026
💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

What version of Codex CLI is running?

0.122.0

What subscription do you have?

ChatGPT Free

Which model were you using?

gpt-5.4 reasoning effort: medium

What platform is your computer?

Microsoft Windows NT 10.0.19045.0 x64

What terminal emulator and version are you using (if applicable)?

command prompt

Codex doctor report

not available (0.122.0 does not support codex doctor)

What issue are you seeing?

Token usage draining approximately 10% per prompt regardless
of task complexity since May 10, 2026.

Observed drain pattern in one session:

  • Before prompt 1: 97%
  • After prompt 1: 89% (-8%)
  • After prompt 2: 79% (-10%)
  • After prompt 3: 67% (-12%)

This did not happen with gpt-5.2 on older versions.

What steps can reproduce the bug?

  1. Open fresh Codex session
  2. Send any prompt (even simple file edit)
  3. Check token usage percentage
  4. Send second prompt
  5. Token drops ~10% each time regardless of task size

What is the expected behavior?

Token usage should be proportional to actual task complexity.
Simple file edits should consume 1-2% not 8-10% per prompt.

Additional information

  • MCP servers: disabled
  • multi_agent: disabled
  • memories: disabled
  • sandbox_mode: danger-full-access
  • Related issues: #22157, #14119, #20796
  • Issue started specifically around May 10, 2026

View original on GitHub ↗

5 Comments

github-actions[bot] contributor · 3 months ago

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

  • #22157
  • #22040
  • #21746
  • #22459

Powered by Codex Action

etraut-openai contributor · 3 months ago

Please update to the latest version of the CLI. You're on a very old version.

Your token usage drain is likely explained by a combination of:

  1. The 2x promotion ended in early May
  2. GPT5.5 is charged at a higher rate than older models
  3. If you have fast mode enabled, it is charged at a 2x rate
sanrishi · 3 months ago

Thank you for the response! @etraut-openai However, I'd like to clarify a few points:

  1. Model: I am using gpt-5.4 not gpt-5.5, so the higher rate for 5.5 doesn't apply here.
  1. Fast mode: I do not have fast mode enabled. My config has model_reasoning_effort = "medium".
  1. 2x promotion: This could explain part of the drain, but not the pattern I'm seeing — approximately 8-10% per single prompt regardless of task complexity. Even a simple file edit costs the same as a complex refactor.
  1. Version: I understand 0.122.0 is old, however upgrading to newer versions introduces the hidden safety reviewer

per-action LLM call (reported in #21746) which makes token usage significantly worse on newer versions.

Could you confirm:

  • Is the 2x promotion ending fully responsible for 8-10% drain per single prompt?
  • Is there a known minimum version that does NOT include the hidden safety reviewer mentioned in #21746?

Thank you!

etraut-openai contributor · 3 months ago

The bug report you linked to that mentioned the "safety reviewer" is factually incorrect. There is no "hidden safety reviewer".

sanrishi · 3 months ago

Hi Eric, one more question as it is confirm on my end that my codex is running at fast at default when i updated the model from 5.2 to 5.4:

I found that service_tier = "fast" was added to my config after I ran /fast accidentally.

According to the docs, fast mode charges at 2x rate.

To permanently disable fast mode, should I:

  1. Just delete the service_tier = "fast" line from config.toml?
  2. Or run /fast off inside the CLI?
  3. Or both?
  4. or another thing to get the fast mode off!

Which is the correct way to ensure fast mode is fully disabled and I am not being charged at 2x rate?

Thank you!