# Codex input vs output tokens, explained for people who post their usage

> Input is what Codex reads, output is what it writes, and cached input is the cheap repeat in between.

URL: https://claudegpt.si/news/codex-input-vs-output-tokens  
Published: 2026-10-26T16:01:01.746Z  
Author: ClaudeGPT Si Newsroom  
Section: Guides  
Topics: Codex, Tokens, OpenAI  
Publisher: ClaudeGPT Si

## Key takeaways

- Codex input tokens are what the model reads and output tokens are what it writes, and OpenAI prices output highest and cached input lowest.
- Agent sessions are mostly input, much of it cached, because Codex resends the chat on every turn.
- UsageAtlas, a free open-source desktop app, shows Codex input, output and cached tokens side by side with estimated cost.

The short answer on Codex input vs output tokens: input tokens are everything the model reads (your prompt, files, chat history and tool output) and output tokens are everything it writes back, including code and plans. OpenAI prices them differently: output costs more per token than input, and cached input costs much less, so the split tells you more about a session than the total does.

If you have ever posted a screenshot of a session with an alarming token count, this is the part that usually explains it.

## The three token types Codex counts

OpenAI's Codex rate table quotes every model in credits per million tokens, in three columns.

| Type | What it is | Relative price in Codex credits |
| --- | --- | --- |
| Input | New text sent to the model: your message, files it opens, command output | Middle |
| Cached input | Input the model has already processed, reused on a later turn | Lowest |
| Output | Text the model writes: code, explanations, tool calls | Highest |

OpenAI also notes that Codex credit billing has no separate charge for writing to the cache. If you use Codex with an API key instead of a ChatGPT plan, API pricing applies, and OpenAI says API token prices should not be used to estimate what a subscription includes.

## Why input dwarfs output in agent work

A coding agent reads far more than it writes. To change ten lines, Codex might open several files, run tests and read their output, and it resends the conversation on every turn. All of that is input. The edit itself is a small amount of output.

Because the history repeats, most of it can be served from cache. That is why a long session can show a huge total that is mostly cached input, while the fresh input and output are a small slice. A big number is not a big bill by itself; look at the split first.

## When output is the expensive part

Output tokens matter most when you ask Codex to write a lot: a whole new module, a long explanation, or a detailed plan. Reasoning effort, which you set with /model on models that support it, can also add output while the model works through a problem. OpenAI's tip here is to match the output to the need: say what format and length you want and separate required work from extras.

## How to see your own split

Inside the Codex CLI, /status shows token usage for the current chat, and /usage daily, weekly or cumulative shows account token activity. Those give you totals in the terminal.

To see input, output and cached tokens side by side over days, [UsageAtlas](https://usageatlas.com) is the simplest option. It is a free and open-source desktop app for Windows, macOS and Linux that reads your local Codex activity and shows tokens, requests, models and estimated cost, with the token types broken out. It does not read your prompts or responses, and cost figures are estimates rather than your bill.

## What to do with the numbers

- **High cached input, low output:** long chats. Start a new one per task with /new, or /compact a long run.
- **High fresh input:** big files, logs or many MCP servers. Point Codex at fewer files and turn off servers you do not need.
- **High output:** long answers or heavy reasoning. Ask for less, or try a lower reasoning effort for routine tasks.

## The short version

Input is what Codex reads, output is what it writes, and cached input is the cheap repeat in between. Output costs the most per token, but input usually makes up most of the count. If you just want today's total, /status and /usage are enough. If you want to know which kind of token is driving your week, [track Codex tokens in UsageAtlas](https://usageatlas.com/guides/view-codex-token-usage).

## Frequently asked questions

### Are Codex output tokens more expensive than input tokens?

Yes. OpenAI's Codex rate table prices output tokens higher than input tokens for each model, and cached input lower than both.

### What are cached input tokens in Codex?

Input the model already processed and reuses on a later turn, such as earlier chat history. OpenAI prices it at a lower credit rate.

### How can I see my Codex input and output tokens?

/status and /usage show token totals in the CLI. UsageAtlas shows input, output and cached tokens by day on your desktop.

## Sources

- [Codex pricing and token rates](https://learn.chatgpt.com/docs/pricing), OpenAI
- [Codex slash commands](https://learn.chatgpt.com/docs/reference/slash-commands), OpenAI
- [Input, output, and cache tokens explained](https://usageatlas.com/guides/cache-tokens-explained), UsageAtlas