OpenAI Codex: What It Is, Price, CLI, Models & Limits

👁 118 reads

Codex is a coding agent from OpenAI: you give it a task, and it reads your repository, edits files, runs commands, and returns a finished change for review. It works in the ChatGPT desktop app, in the terminal (Codex CLI), in the cloud, and as an IDE extension, and you get access as part of your ChatGPT plan - starting from ChatGPT Plus, with a limited version even in the free plan.

What Codex Is (and How It Differs from the Old 2021 Codex)

Today’s Codex is an agent, not a model. It doesn’t suggest individual lines, but accepts tasks like “fix the tests,” “add Google login,” or “review this pull request,” plans steps, works on files and in the terminal, and finally shows code differences (diff) for approval. Under the hood, models from the GPT-6 family work, which you choose yourself.

The name has a history, though, which creates confusion in search results:

• Codex from 2021 - a language model announced on August 10, 2021, essentially GPT-3 fine-tuned on code from GitHub. It powered the first version of GitHub Copilot. Today it’s a closed chapter and has nothing in common with the current product except the name.
• Codex CLI - an open-source agent in the terminal, released on April 16, 2025 under the Apache-2.0 license.
• Codex in the cloud - a research version from May 16, 2025, initially on the codex-1 model.
• Codex app - a separate desktop application from February 2026, merged with the ChatGPT app for Mac and Windows on July 9, 2026.

So if you came here asking “what is codex,” the answer for October 2026 is: an OpenAI agent for working with code, included in ChatGPT plans. Our older description of the agent work concept is in the article Codex: agent OpenAI, któremu oddaje się zadanie, a nie linijkę kodu.

Where It Works: App, CLI, Cloud, IDE

Codex is one agent with multiple entry points. The usage limit is shared, so you don’t pay separately for terminal and separately for the app.

In ChatGPT you’ll also see the name ChatGPT Work - this is the agent mode in ChatGPT itself. According to OpenAI’s pricing, ChatGPT Work and Codex use the same limit and the same credits.

How to Get Started Step by Step

The simplest path is the ChatGPT desktop app, where Codex is one of the views. If you prefer the terminal, you can install CLI in a minute (we described Windows installation with typical errors in the guide Codex CLI na Windows):

1. Install CLI using one of these methods:
   • npm (requires Node.js): npm install -g @openai/codex
   • Homebrew on macOS: brew install --cask codex
   • installer on macOS and Linux: curl -fsSL https://chatgpt.com/codex/install.sh | sh
   • installer on Windows (PowerShell): powershell -ExecutionPolicy ByPass -c "irm https://chatgpt.com/codex/install.ps1 | iex"
2. Check version: codex --version. Later you update with codex update, and diagnose installation problems with codex doctor.
3. Log in with ChatGPT account: enter a project directory, type codex and choose Sign in with ChatGPT. You can check login status with codex login status. Alternative is API key (codex login --with-api-key) - then you pay for tokens and lose cloud features.
4. Make a commit before the first task to have a rollback point.
5. Give a small task, e.g., “Tell me about this project.” Useful commands in session: /init creates an AGENTS.md file with instructions for the agent, /model changes model and reasoning level, /permissions sets what the agent can do on its own, /review starts change review, /status shows remaining limit.

For automation, use codex exec "command" - the agent works without interaction and finishes with a report. The codex app command opens the desktop app (or its installer), and codex cloud lets you browse cloud tasks and apply their results to your local repository.

Our CLI test from October 3, 2026. We prepared a mini Python project with a function formatting amounts and a test that wasn’t passing. We ran codex exec -m gpt-6.1-sol -s workspace-write with a command in Polish: fix the function, don’t touch tests, describe the cause. Result:

• time: 87 seconds, usage: 7636 tokens;
• fix was accurate: the cause was truncating cents through int() with floating-point imprecision, and the test passes after the fix;
• it honestly noted that in the sandbox it couldn’t run pytest (missing package, blocked access to PyPI), so it checked the test directly with Python;
• our note: for monetary amounts we’d prefer the Decimal type, and the agent chose regular number formatting. It works, but with money it’s worth mentioning this explicitly in the command.

Models in Codex

You choose the model in the app from a list, and in CLI with the /model command or -m flag. Prices in the table are API rates (apply when working with API key), not what you’ll pay in subscription.

GPT-6 Astra was released on September 3, 2026. GPT-6.1 Sol was shown at DevDay on September 29 - OpenAI claims that in the DeepSWE v1.1 benchmark it matches Astra at roughly one-fifth the cost. This is the manufacturer’s claim, not our measurement. Astra’s successor, GPT-6.1 Astra, was halted before launch and is not available in Codex.

There are also speed modes: Fast consumes subscription limit 2.5 times faster, and Astra Ultrafast (up to 8 times faster generation, about 300 tokens per second according to OpenAI) 8 times faster and is only in Pro 500 plan and in some Enterprise and Edu plans. Details in the article about Ultrafast mode.

How Much It Costs - ChatGPT Plans and Codex Limits

Codex doesn’t have a separate subscription. You pay for a ChatGPT plan, and Codex is included. Monthly prices:

Full plan comparison, also beyond Codex, is in the article ChatGPT cena w Polsce, and we write about the most expensive plan in ChatGPT Pro 500. Business plan costs according to OpenAI pricing $20 per person with annual payment and $25 with monthly (minimum 2 people).

Limits in Plus plan. OpenAI provides estimated ranges of messages executed locally per 5 hours:

• GPT-6 Astra: 5-45 messages;
• GPT-6.1 Sol: 15-160;
• GPT-6 Sol: 15-150;
• GPT-6 Luna: 350-3000.

Ranges are wide because usage depends on repository size, session length, and number of tools. Cloud tasks consume limit faster than local ones, image generation 3-5 times faster, and each connected MCP server adds context. Beyond the 5-hour limit, weekly limits may apply. Pro plans according to documentation don’t currently have a 5-hour limit. After exhausting the limit, Plus and Pro can purchase credits instead of changing plans.

Note for Pro 200 holders: those who had an active subscription before the change keep their previous limit until October 29, 2026. Then they move to a lower allocation at the same price. How much limit individual models consume and how to check it, we detailed in the article Codex limity w Plus.

Codex vs Claude Code vs Cursor vs Antigravity

All four do the same thing in outline: the agent reads the project, changes files, and runs commands. They differ in whose models you get and in what form.

The last row is our opinion, not a ranking. We work in parallel in Codex and Claude Code, so we described a detailed comparison of these two separately in the article Codex czy Claude Code. A broader tool overview is in AI do programowania.

Our Use Cases

Local Codex history on our Windows computer contains about 2,400 sessions since May 2026, of which all but two we ran in the desktop app, not in terminal. In the last 400 sessions, GPT-6 Astra worked most often (220 times), then GPT-5.6 Sol (85), and since September 29 we’re moving some tasks to GPT-6.1 Sol. What we use it for:

• Research with price comparison. E.g., “find the cheapest API providers for video models that must work in Poland, and compare prices.” Codex searches the web, collects a table, and immediately saves it to a file in the task folder.
• Landing pages and prototypes from scratch. The app creates a separate folder with date for each thread, so a quick page prototype or startup mockup doesn’t mix with other projects.
• Screenshot as command. Most of our tasks start with a pasted screenshot: error, someone else’s page, pricing panel. This is the most convenient difference compared to working in terminal.
• Finding old work. Commands like “find that hosting project we did” usually work because files from previous threads are in folders on disk that the agent has access to.
• Series of graphics. 8 logo proposals or 12 product photos in 4:5 format with file names specified in the command. Here we watch the limit because images consume it faster.
• Checking if a site is safe. We paste an address or screenshot and ask for assessment before clicking or downloading anything.
• Permanent rules in AGENTS.md. In the global ~/.codex/AGENTS.md file we keep rules that should always apply, e.g., that in texts we only use plain hyphens. This way we don’t repeat them in every command.

What we don’t do: we don’t give the agent full permissions in directories with client data and passwords, and before a major change we make a commit. We wrote about risks with company code in the article agent do kodu w firmie.

What Codex Got at DevDay 2026

At OpenAI DevDay 2026 (September 29), five announcements concerned Codex:

• Codex in the cloud with reusable environments: you configure repository and dependencies once, and subsequent tasks start from ready configuration, even with laptop closed. Step-by-step description: Codex w chmurze.
• Refreshed CLI: voice control, /agents view to track multiple tasks at once, better session resumption and work with git worktree.
• Code review in ChatGPT app with automatic pull request reviews on GitHub (GitLab in preview).
• Codex Security Cloud for scanning repositories on demand or on schedule, in Pro, Business, Enterprise, and Edu.
• GPT-6.1 Sol and Astra Ultrafast as new model and speed options.

Separately, OpenAI announced “Sign in with ChatGPT,” meaning using plan limits in partner tools - we described this in the article Zaloguj się przez ChatGPT.

Frequently Asked Questions

Codex - What Is It?

It’s OpenAI’s coding agent. It receives a task, reads code on its own, edits files, runs commands, and returns changes for review. It works in the ChatGPT desktop app, in terminal, in cloud, and in code editors. Don’t confuse it with the Codex model from 2021 that powered the first GitHub Copilot.

How Much Does Codex Cost?

Codex is included in your ChatGPT plan. Full access starts from ChatGPT Plus at $20 per month. Pro costs $100 (Pro 100), $200 (Pro 200), and $500 (Pro 500). With API key you pay for tokens, e.g., $2 per million input tokens in GPT-6.1 Sol.

Is Codex Free?

Partially. In Free and Go plans, OpenAI provides GPT-6 Luna in the desktop app, rolling out gradually. Codex CLI itself is open-source and you can connect it to local models through Ollama or LM Studio (--oss flag), but quality then depends on the model on your computer.

How to Install Codex CLI?

Easiest via npm: npm install -g @openai/codex, on macOS also brew install --cask codex. Then in a project directory you type codex and log in with ChatGPT account. You check version with codex --version, as of October 3, 2026 current is 0.160.0. We described Windows installation with typical errors in the guide Codex CLI na Windows krok po kroku.

Codex or Claude Code - What to Choose?

If you already have ChatGPT Plus, start with Codex because you don’t pay extra. Claude Code requires a paid Claude plan (from $20) and gives you Anthropic models. We use both: Codex mainly in the app with multiple threads and screenshots, Claude Code for long terminal sessions. Details in the article Codex czy Claude Code.

Does Codex Work in Polish?

Yes, it understands commands in Polish without problems and responds in Polish - in our CLI test it described the error cause in Polish without an additional request. The CLI interface and system messages are in English.

Which Model to Choose in Codex?

For daily work GPT-6.1 Sol, which is default in CLI and has several times higher limit in Plus plan than Astra. We save GPT-6 Astra for the most difficult tasks, and GPT-6 Luna for simple, repeatable actions.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources and Check Date

• OpenAI: DevDay 2026 Recap (09/29/2026)
• OpenAI: Codex / ChatGPT Work pricing and limits
• OpenAI: Codex CLI documentation
• OpenAI: Codex IDE extension
• OpenAI: Codex in the cloud
• GitHub: openai/codex repository (installation, license)
• npm: @openai/codex package (version 0.160.0)
• Releasebot: Codex release history
• Wikipedia: OpenAI Codex (AI agent) and OpenAI Codex (language model) - historical dates
• Anthropic: Claude pricing, Cursor pricing, Antigravity pricing

API prices and competitor plan prices are in USD. Limits are OpenAI estimates, not guaranteed numbers. CLI test and session statistics are our measurements on our own computer. Checked October 3, 2026.