Why Is My Codex CLI Token Usage Suddenly So High? | Diagnosis & Fixes

Mar 2, 2026

Problem

My Codex CLI token usage spiked from ~2% per hour to 9% in two hours. and I’m not alone.

I checked my Reddit and and found I wasn’t crazy. Multiple users reported the same thing:

> My codex tokens suddenly have been exhausted over the past week - has something changed?
> > 8% daily quota for 20 lines of code changed
> > 25% weekly limit burned in 30 minutes

Environment

Codex CLI 0.106.0
ChatGPT Pro plan ($20/month)
macOS,14.x, Sonoma 14.7.2

What Happened

I was working on a feature implementation when I noticed my weekly usage jumping from 62% to 59% to 62% repeatedly. That bouncing pattern was suspicious.

I decided to analyze my token consumption. Here’s what I found

User: killtheperfect
Date: 2026-02-27
Sessions: 62
Token-count events: 4,499

The output shocked me:

Median context per turn: ~96k tokens
p95 context per turn: ~200k tokens
Baseline startup overhead: 21-22k tokens (up from 12-15k)
Shell tool outputs: 90.3% of all tool-output characters

The Root Cause

I discovered three culprits:

1. Shell Tool Output Explosion

┌──────────────────────────────────────────────────────────────┐
│              User Request                               │
└──────────────────────────────────────────────────────────────┘
                          │
                          ▼
              ┌────────────────────────┐
              │ System Prompt (21k)   │
              └────────────────────────┘
                          │
                          ▼
    ┌─────────────────────────────────────────────────────────────┐
    │  Shell Commands (rg, git diff, cat logs)           │
    │  90.3% of context window                              │
    └─────────────────────────────────────────────────────────────┘
                          │
                          ▼
                  ┌─────────────────┐
                  │ Token Exhaustion │
                  └─────────────────┘

Every rg search, git diff, and test output got dumped into my context window.

2. Elevated Baseline Overhead

Previous: 12-15k tokens per session startup
Current: 21-22k tokens per session startup
Increase: ~50-60% more overhead before any real work

This started around version 0.106.0 based on community reports.

3. Context Accumulation

Session starts with 21k overhead
First shell command adds 10-50k tokens
Context compaction hasn’t occurred yet
By turn 5-10, I’m at 100k+ tokens consistently

How to Diagnose

I tried Codex’s built-in diagnostic tools.

Check Rate Limits:

{ "method": "account/rateLimits/read", "id": 1 }

Response:

{
  "usedPercent": 87,
  "windowDurationMins": 10080,
  "resetsAt": "2026-03-03T00:00:00Z"
}

Monitor Real-Time Usage:

{
  "method": "thread/tokenUsage/updated",
  "params": {
    "threadId": "thr_123",
    "usage": {
    "totalTokens": 96000,
    "promptTokens": 94000,
    "completionTokens": 2000
  }
}

The promptTokens field showed my context window was bloated with shell output.

Solutions

I tried several approaches to reduce consumption.

Attempt 1: Limit Shell Output (Partial Success)

# Before
rg "authentication" .
# After
rg "authentication" --max-count=10 -l

This helped, but I was still hitting high usage because of accumulated context.

Attempt 2: Manual Compaction (Better)

{ "method": "thread/compact/start", "id": 25, "params": { "threadId": "thr_b" } }

This reduced context by ~40%, but I had to trigger it manually.

Attempt 3: Switch to Mini Model (Best Results)

For routine tasks like refactoring and documentation updates, I switched to GPT-5.1-Codex-Mini:

Result: ~4x reduction in token consumption for simple tasks.

What Actually Worked

After experimenting, I found these strategies reduced my consumption by 60-80%

1. Targeted Shell Commands

# Bad: Full diff
git diff main...feature

# Good: Summary first
git diff --stat main...feature | head -20
git diff main...feature -- src/auth/ | head -100

2. Manual Compaction Trigger compaction after large outputs:

{ "method": "thread/compact/start", "id": 25, "params": { "threadId": "thr_current" } }

3. Use Mini for Routine Tasks Switch to GPT-5.1-Codex-Mini for:

Simple refactoring
Documentation updates
Small bug fixes

4. Disable Unused MCP Servers Each MCP server adds context. Disable when not needed.

5. Precise Prompts

Bad: “Help me fix the authentication issue”
Good: “Fix the JWT validation error in /src/auth/jwt.ts line 45. The error is ‘invalid signature’.”

Quick Reference Checklist

Before Session:

Disable unused MCP servers
Review AGENTS.md for redundancy
Choose appropriate model (Mini for routine tasks)

During Session:

Use targeted shell commands with output limits
Monitor token usage via notifications
Trigger compaction after large outputs

After Session:

Check rate limit status
Review which commands consumed most tokens
Consider forking threads for new tasks

Summary

In this post, I diagnosed why my Codex CLI token usage suddenly spiked. The key points are:

Shell tool outputs consume 90%+ of context
Baseline overhead increased 50-60% (12-15k to 21-22k tokens)
Proactive management reduces usage by 60-80%

Start with account/rateLimits/read to assess your current state, then apply targeted shell commands and manual compaction to keep your context window lean.

Final Words + More Resources

My intention with this article was to help others share my knowledge and experience. If you want to contact me, you can contact by email: Email me

Here are also the most important links from this article along with some further resources that will help you in this scope:

Oh, and if you found these resources useful, don’t forget to support me by starring the repo on GitHub!