OpenAI Codex in 2026: features, plans, and practical trade-offs

This product overview explains where Codex runs, what its plans include, and which limits and permission settings to check in your own project. It uses the official product pages cited below, with prices and model details recorded as of October 4, 2026. It is not a dated collection of user reviews or a measured survey of satisfaction.

First, let me clarify what we’re talking about, because this is a common source of confusion. Codex is an AI coding agent from the same OpenAI that makes ChatGPT. It digs through your files, edits them, and runs commands, while you watch and say “yes” or “try again.” If you first want to understand how an agent itself works, there’s a separate explainer, what an AI agent is, in plain language, and how to install Codex from scratch is shown in what it is and how to launch it. This article focuses on practical product choices rather than installation or a review survey.

Ailoot documentation: API, keys, and setup.

What is Codex, and why does the version matter?#

OpenAI has reused the name Codex before, and that’s the first source of confusion: the old Codex from 2021 was a code-completion model, and support for it ended in March 2023. The current Codex is a separate product, an agent: it edits project files and runs commands on its own.

The Codex CLI open-source code is on GitHub under the Apache-2.0 license: as of October 4, 2026, the repository has 127,810 stars and 20,010 forks (GitHub API data, checked for this article). The project changes actively, which explains the second source of variation in reviews: over the past year, Codex has gone through several model changes—from GPT-5-Codex to the GPT-5.6 family, and in fall 2026 to GPT-6. A comparison of one version may not describe another. The product snapshots below are dated where relevant; there are no attributed user quotations or issue case studies in this overview.

Which Codex features are useful to evaluate?#

The inclusion in ChatGPT plans lowers the barrier to trying the product. Codex isn’t sold as a separate subscription: it’s included in every ChatGPT plan, from the free Free plan to Pro.

Another practical feature is the range of places where it can run.

Which limits and permission settings should you check?#

Inspect the approval and sandbox settings before the pilot. Record the exact command and permission prompt if behavior differs from what you expected, rather than assuming a generic issue applies to your setup.

Check remaining usage and reset windows before a long task; a low entry price does not make usage unlimited.

On complex tasks, define a small working result and run its tests before expanding scope.

Figuring out what goes into the agent’s context so it stays on task and doesn’t lose the thread is exactly what’s called context engineering. The tool, whether Codex or Claude, is secondary: what makes the biggest difference to the result is what the agent is given as context.

Codex prices and plans for October 2026#

Here is the dated plan snapshot from OpenAI’s official pricing page (developers.openai.com/codex/pricing, checked 04.10.2026):

The billing model changed on April 2, 2026: OpenAI dropped the old “message” limits and switched to credits tied to tokens, as in the API. Current rates from the official pricing page: GPT-6 Astra—250 credits per 1M input tokens and 1,250 per output tokens; GPT-6.1 Sol—50 and 250; GPT-6 Luna—2.5 and 12.5. These are Standard rates; purchased credits and Fast mode cost more (Fast—2x, Astra Ultrafast—6x the standard rate).

Practical takeaway: if you already pay for ChatGPT Plus for chat, Codex comes at no extra charge. You can also start with Free to try it out. Consider switching to Pro when Plus limits start getting in the way of your daily work.

Models currently available in Codex#

This is the most noticeable change since summer 2026: Codex moved from the GPT-5.6 family to GPT-6. According to the official Codex model documentation (developers.openai.com/codex/models, checked 04.10.2026), three models are recommended now:

• GPT-6 Astra—the most capable model for complex work: coding, applications, research, and extended reasoning.
• GPT-6.1 Sol—a model close to Astra in quality for repeatable, long-running work, at a lower credit cost. It’s the default in fresh CLI installations (model: gpt-6.1-sol medium, according to the official terminal screentest on the same page).
• GPT-6 Luna—the most economical model for clearly defined, high-volume tasks: data extraction, classification, and structured summaries.

The model page also says that GPT-5.6 Sol, Terra, and Luna “remain available during the rollout”—the old generation was given time to transition instead of being shut off abruptly. Older models, however, are being retired on a schedule: GPT-5.4 and GPT-5.4-mini were turned off in Codex on August 31, 2026; GPT-5.3-Codex-Spark on September 14, 2026; and GPT-5.5 is scheduled to go on October 14, 2026, according to the same documentation. If one of these models is still in a saved configuration, it should be replaced with an available one; otherwise, tasks will start failing with an error after the shutdown date.

You can switch the model and reasoning depth on the fly with the /model command in the terminal. The advice still holds: don’t keep the maximum reasoning level on for simple edits—it uses credits without improving quality. Heavy mode makes sense for a heavy task.

Codex CLI, editor extension, or cloud—where should you run it?#

Codex has several ways into the same product, so there’s plenty to choose from.

Codex CLI runs in the terminal. It’s installed with one command, which is current as of October 2026 according to the official CLI page (developers.openai.com/codex/cli):

npm install -g @openai/codex

Next, run codex in the project folder. It’s a good idea to start with /init: Codex creates an AGENTS.md file with instructions about the project, then uses it so it doesn’t have to guess from scratch each time.

The editor extension puts Codex in the VS Code window. For most people, this is the easiest way in: the code is in front of you, the agent is alongside it, and everything is in one place.

The cloud version runs in a browser inside ChatGPT. It’s useful for long tasks: you hand off the task, close the tab, and Codex works on its own on OpenAI’s servers; you pick up the result later. There’s a desktop app for macOS and Windows, and a phone app for checking a task’s status from the couch.

Who should try Codex, and what should a pilot check?#

If you already pay for ChatGPT, start with a small task using the included access. Choose one observable result: a script that handles a real input, a bug reproduced and fixed by a test, or a pull-request finding you can verify.

Check the command approval behavior and usage consumed during that pilot. For an agent comparison, use the same starting commit and acceptance criteria; see Codex vs Claude Code and migrating from Claude Code to Codex.

If you need access to AI models through Ailoot, explore Ailoot.