Track Codex Tokens with Absolute Numbers

OpenAI Codex tells you "70 % consumed." It does not tell you how many tokens that is. 70 % could be 300,000 tokens or 3,000,000 tokens. The difference matters when you are budgeting a multi-file refactor or comparing Codex against Claude Code on the same task.

This guide shows how to track Codex tokens with real numbers — input, output, and cost — not just percentages.

1. Why Codex Hides the Number

Codex's /status command shows subscription tier usage as a percentage. This is by design:

• It updates in real time
• It works across all models in your tier
• It requires no additional infrastructure

But it has hard limits:

• No absolute token count
• No input vs. output split
• No cross-session aggregation
• No tool comparison

2. Where Codex Writes the Real Data

Codex logs every interaction to local JSONL files:

~/.codex/sessions/**/*.jsonl

Each event contains:

• model name, depending on your Codex version and config
• input tokens
• output tokens
• reasoning tokens (if applicable)
• timestamp

3. How to Read the Logs Manually

One event looks like this:

{
  "type": "event_msg",
  "payload": {
    "type": "token_count",
    "info": {
      "last_token_usage": {
        "input_tokens": 4523,
        "cached_input_tokens": 3200,
        "output_tokens": 891,
        "reasoning_output_tokens": 120,
        "total_tokens": 5534
      }
    }
  },
  "timestamp": "2026-05-20T10:30:00Z"
}

To get a daily total, you would:

1. Find all *.jsonl files under ~/.codex/sessions/
2. Parse each JSON line
3. Sum input_tokens and output_tokens by date
4. Multiply by the model's per-token cost

This is tedious. A token tracker automates it.

4. What TokenBBQ Shows You

TokenBBQ reads Codex's session JSONL files automatically and serves:

• Absolute daily token totals (input + output)
• Cost in USD (fetched from LiteLLM pricing)
• Cross-session aggregation (all Codex sessions combined)
• Multi-tool comparison (Codex vs. Claude vs. Gemini)
• Live dashboard at localhost:3000

5. Codex vs. Claude Code: Token Profiles

• Codex records token-count snapshots in local session JSONL files
• Codex can chain multiple tool calls, so long agent runs may accumulate usage quickly
• Codex exposes controls such as --model, --ask-for-approval, and --sandbox
• Claude Code records usage in its own project JSONL logs, including cache read/write fields when available
• TokenBBQ normalizes both tools into one daily, model, and project view

6. Quick Start

npx tokenbbq@latest

No install. No config. No API key. Dashboard opens in browser.

7. Related Guides

• How to Track OpenAI Codex Token Usage — the full guide with dashboard setup and Pill overview.
• AI Token Tracker — Compare All Major Coding Tools — multi-tool comparison in one dashboard.
• How to Track Claude Code Token Usage — the Anthropic side with window tracking.
• Track AI Token Usage Per Project — split costs across repositories.