Why Codex tokens per session keep climbing, and how to watch them
Each turn resends the chat, so long sessions grow fast. Watch the live count in /status and the history in a desktop tracker.
By ClaudeGPT Si Newsroom, Editorial team3 min read
Codex tokens per session grow faster than the work does, because each turn sends the whole chat back to the model. You can watch the count live with /status in the Codex CLI, and keep the totals afterwards in UsageAtlas, a free desktop tracker that turns your local Codex activity into a daily history.
There is no normal number of tokens for a Codex session. A short fix and a two-hour refactor can differ by orders of magnitude, so the useful question is what drives your own sessions up.
What goes into a session's token count
OpenAI's pricing page lists what uses tokens in Codex: your prompt, files, chat history, tool results and the reply. In a coding session, that adds up in a few predictable ways.
| Driver | Why it grows the count | What to watch |
|---|---|---|
| Chat history | Resent on every turn | Long chats with many small follow-ups |
| Files read | Each file is input | Whole directories, logs, generated files |
| Tool results | Command output comes back as input | Test runs and builds with long output |
| Instructions | AGENTS.md is added to the prompt | A long root AGENTS.md |
| MCP servers | Each adds context to messages | Servers this project does not use |
| Output | Code, plans and replies | Asking for more than you need |
The first row is the one that surprises people. Turn twenty carries turns one to nineteen with it. Much of that repeat shows up as cached input, which OpenAI prices at a much lower credit rate than fresh input, but it still counts.
Watching a session live
Inside Codex, /status shows token usage and remaining context capacity for the current chat. The footer shows how much context is left as a percentage, and /statusline lets you add token and limit fields so you do not have to ask. When the context gets close to full, Codex compacts the history on its own, or you can run /compact.
That covers the session you are in. It does not tell you how this session compares with yesterday's, or which project eats most of your week.
Watching sessions over time
This is where a tracker helps. UsageAtlas reads Codex activity already on your computer and shows tokens, requests, models and estimated cost by day. It splits input, output and cached tokens, groups usage by project folder, and its Today view compares the current day with your recent baseline. It runs on Windows, macOS and Linux, is free and open source, and needs no account for local tracking.
UsageAtlas also shows the Codex session and weekly windows and reset times where Codex reports them, and can send one desktop alert per window at a threshold you pick. Cost figures are estimates; your OpenAI account is the bill.
Patterns worth spotting
After a week or two of history, a few patterns tend to stand out.
- Marathon chats. One chat that runs all afternoon usually costs more than several short ones for the same work. /new between tasks resets the history.
- One noisy project. A repository with huge test output or a long AGENTS.md can dominate the totals. The project view shows this quickly.
- Model choice. OpenAI's rate table prices models very differently per token. If one model takes most of your estimated cost, try a smaller one for routine tasks.
- Cache-heavy days. A large total with a high cache share is usually long context, not a lot of new work.
When /status is enough
If you run a few short sessions a day and never hit a limit, the live count in /status tells you all you need. Tracking sessions over time is worth it when limits arrive earlier than you expect, or when you want to know which habit to change.
To see your own Codex tokens per session add up over a week, install the free UsageAtlas app and check back after a few days of normal work.
FAQ
How many tokens does a Codex session use?
There is no typical number. It depends on how long the chat runs, how many files and tool results it reads, and the model. Measure your own sessions.
How do I see tokens for the current Codex session?
Type /status in the Codex CLI. It shows token usage and remaining context capacity for the chat you are in.
Why is most of my Codex token count cached input?
Each turn resends the conversation, and much of that repeat is served from cache, which OpenAI prices lower than fresh input.
Originally published on claudegpt.si