In partnership with
AI Spotlight · Tool Deep Dive
ChatGPT Codex Plan: The Most Underrated Upgrade Right Now
Five million developers are already using it. The Codex plan inside ChatGPT is being called the best bang-for-buck AI subscription of June 2026, and most people still have not figured out what it actually does.
There is a quiet argument happening in developer communities right now. On one side: people who tried Codex early, found it limited, and moved on to Claude Code or Cursor. On the other side: people who have been watching the Codex changelog since April 2026 and are increasingly convinced that the conversation needs updating. A lot has changed. The product that exists today is not the product that existed six months ago, and the gap between perception and reality has become large enough to be worth addressing directly.
Codex grew from three million to five million weekly active developers between April and June 2026. Goal mode shipped stable. GPT-5.5 arrived as the default model. Browser use became a real feature. Sites launched for building and deploying apps entirely within Codex. The product has been shipping at a velocity that most coverage has not kept pace with. This issue gives you the current picture: what the plans actually include, what Codex genuinely does well, who should upgrade, and where the real limitations still sit.
* * *
Start here
What Codex actually is and how it differs from ChatGPT
Codex is not a chatbot. That distinction matters more than it sounds. While ChatGPT gives you answers in a conversation window, Codex operates as a cloud-based software engineering agent. You connect your GitHub repository, give it a task, and it works asynchronously in an isolated cloud sandbox: reading your codebase, writing code, running tests, committing changes to a branch, and returning a diff with terminal logs and test results as verifiable evidence of what it did. It does not just suggest code. It acts on your codebase.
The current default model is GPT-5.5, which scored 82.7% on Terminal-Bench 2.0, up from 75.1% for the previous GPT-5.4. One developer described it as "the first coding model I have used that has serious conceptual clarity." Each task runs in its own sandboxed container preloaded with your repository and configured dependencies, and you can queue multiple tasks to run in parallel. While Codex is working on one thing, you can queue three more and review them all when you come back.
|
Codex vs ChatGPT: the one-line difference ChatGPT suggests code in a conversation. Codex runs tasks on your actual repository in a sandboxed cloud environment, commits the result to a branch, and hands you a diff to review. One is an advisor. The other is a contractor who does the work and shows you receipts. |
* * *
The plans explained
What each plan actually includes in June 2026
Codex is not a separate product you buy. It is bundled into every ChatGPT plan tier. The distinction between plans is how much Codex you get and which features unlock. Here is the current picture as of June 2026, with the detail that actually matters for decision-making.
|
|
|
|
One pricing detail that caught a lot of developers by surprise: since April 2026, Codex switched from per-message counting to token-based credits on a rolling 5-hour window. This means the same plan can feel very different depending on task complexity. A simple bug fix burns far fewer credits than a long-horizon refactor. Community estimates for power users land at $100 to $200 per developer per month in API-equivalent cost, a figure worth knowing before you build Codex deeply into a team workflow.
* * *
Where it shines
The four things Codex does better than any alternative right now
Codex is not the best tool for every coding task. But there are four specific use cases where its architecture, cloud-based, asynchronous, parallel, and sandbox-isolated, makes it meaningfully better than alternatives. These are the cases that have driven the user growth from three to five million developers in a single quarter.
|
Where Codex has a structural edge Batch maintenance across large codebases: Renaming patterns, updating deprecated API calls, standardising error handling, migrating configuration formats. Tasks where the pattern is clear, the work is repetitive, and the risk of interrupting your own flow state is high. Codex handles these in the background while you work on something else. OpenAI published a case study in June 2026 of an internal product built with roughly 1,500 pull requests over five months from a team of three to seven engineers, with zero manually written code. Codex did the implementation; the humans did the review. Parallel task execution: The ability to queue four or five tasks simultaneously is one of the genuine differentiators that no chat-based coding tool can match. While Codex works on the bug fix, you queue the refactor, the documentation update, and the test coverage expansion. When you come back from a meeting, there are four branches waiting for your review rather than four tasks waiting for you to start them. Bug fixing with a failing test: When there is a failing test and a clear bug description, Codex's approach of making the test pass without breaking anything else works reliably. The sandbox runs the tests. You see the test results alongside the code changes. It is a fundamentally more verifiable output than a chat-based suggestion you have to manually test yourself. Building and deploying internal tools with Sites: The June 2 Sites launch allows Codex to build a full interactive web application from a prompt, deploy it to OpenAI-managed hosting, and return a shareable URL with workplace authentication through Sign in with ChatGPT. Sales dashboards, project trackers, internal knowledge bases. Work that previously required a junior developer for two weeks now takes one Codex session. This feature is currently available for Business and Enterprise plans but broader rollout is expected. |
The common thread across all four: Codex is most powerful when the task is well-defined, the output is verifiable, and you do not want to stop what you are currently doing in order to do it. It is not a replacement for thinking. It is a way to execute while you think.
AI Agents Are Reading Your Docs. Are You Ready?
Last month, 48% of visitors to documentation sites across Mintlify were AI agents, not humans.
Claude Code, Cursor, and other coding agents are becoming the actual customers reading your docs. And they read everything.
This changes what good documentation means. Humans skim and forgive gaps. Agents methodically check every endpoint, read every guide, and compare you against alternatives with zero fatigue.
Your docs aren't just helping users anymore. They're your product's first interview with the machines deciding whether to recommend you.
That means: clear schema markup so agents can parse your content, real benchmarks instead of marketing fluff, open endpoints agents can actually test, and honest comparisons that emphasize strengths without hype.
Mintlify powers documentation for over 20,000 companies, reaching 100M+ people every year. We just raised a $45M Series B led by @a16z and @SalesforceVC to build the knowledge layer for the agent era.
AI Spotlight · Tool Deep Dive
ChatGPT Codex Plan: The Most Underrated Upgrade Right Now
Five million developers are already using it. The Codex plan inside ChatGPT is being called the best bang-for-buck AI subscription of June 2026, and most people still have not figured out what it actually does.
There is a quiet argument happening in developer communities right now. On one side: people who tried Codex early, found it limited, and moved on to Claude Code or Cursor. On the other side: people who have been watching the Codex changelog since April 2026 and are increasingly convinced that the conversation needs updating. A lot has changed. The product that exists today is not the product that existed six months ago, and the gap between perception and reality has become large enough to be worth addressing directly.
Codex grew from three million to five million weekly active developers between April and June 2026. Goal mode shipped stable. GPT-5.5 arrived as the default model. Browser use became a real feature. Sites launched for building and deploying apps entirely within Codex. The product has been shipping at a velocity that most coverage has not kept pace with. This issue gives you the current picture: what the plans actually include, what Codex genuinely does well, who should upgrade, and where the real limitations still sit.
* * *
Start here
What Codex actually is and how it differs from ChatGPT
Codex is not a chatbot. That distinction matters more than it sounds. While ChatGPT gives you answers in a conversation window, Codex operates as a cloud-based software engineering agent. You connect your GitHub repository, give it a task, and it works asynchronously in an isolated cloud sandbox: reading your codebase, writing code, running tests, committing changes to a branch, and returning a diff with terminal logs and test results as verifiable evidence of what it did. It does not just suggest code. It acts on your codebase.
The current default model is GPT-5.5, which scored 82.7% on Terminal-Bench 2.0, up from 75.1% for the previous GPT-5.4. One developer described it as "the first coding model I have used that has serious conceptual clarity." Each task runs in its own sandboxed container preloaded with your repository and configured dependencies, and you can queue multiple tasks to run in parallel. While Codex is working on one thing, you can queue three more and review them all when you come back.
|
Codex vs ChatGPT: the one-line difference ChatGPT suggests code in a conversation. Codex runs tasks on your actual repository in a sandboxed cloud environment, commits the result to a branch, and hands you a diff to review. One is an advisor. The other is a contractor who does the work and shows you receipts. |
* * *
The plans explained
What each plan actually includes in June 2026
Codex is not a separate product you buy. It is bundled into every ChatGPT plan tier. The distinction between plans is how much Codex you get and which features unlock. Here is the current picture as of June 2026, with the detail that actually matters for decision-making.
|
|
|
|
One pricing detail that caught a lot of developers by surprise: since April 2026, Codex switched from per-message counting to token-based credits on a rolling 5-hour window. This means the same plan can feel very different depending on task complexity. A simple bug fix burns far fewer credits than a long-horizon refactor. Community estimates for power users land at $100 to $200 per developer per month in API-equivalent cost, a figure worth knowing before you build Codex deeply into a team workflow.
* * *
Where it shines
The four things Codex does better than any alternative right now
Codex is not the best tool for every coding task. But there are four specific use cases where its architecture, cloud-based, asynchronous, parallel, and sandbox-isolated, makes it meaningfully better than alternatives. These are the cases that have driven the user growth from three to five million developers in a single quarter.
|
Where Codex has a structural edge Batch maintenance across large codebases: Renaming patterns, updating deprecated API calls, standardising error handling, migrating configuration formats. Tasks where the pattern is clear, the work is repetitive, and the risk of interrupting your own flow state is high. Codex handles these in the background while you work on something else. OpenAI published a case study in June 2026 of an internal product built with roughly 1,500 pull requests over five months from a team of three to seven engineers, with zero manually written code. Codex did the implementation; the humans did the review. Parallel task execution: The ability to queue four or five tasks simultaneously is one of the genuine differentiators that no chat-based coding tool can match. While Codex works on the bug fix, you queue the refactor, the documentation update, and the test coverage expansion. When you come back from a meeting, there are four branches waiting for your review rather than four tasks waiting for you to start them. Bug fixing with a failing test: When there is a failing test and a clear bug description, Codex's approach of making the test pass without breaking anything else works reliably. The sandbox runs the tests. You see the test results alongside the code changes. It is a fundamentally more verifiable output than a chat-based suggestion you have to manually test yourself. Building and deploying internal tools with Sites: The June 2 Sites launch allows Codex to build a full interactive web application from a prompt, deploy it to OpenAI-managed hosting, and return a shareable URL with workplace authentication through Sign in with ChatGPT. Sales dashboards, project trackers, internal knowledge bases. Work that previously required a junior developer for two weeks now takes one Codex session. This feature is currently available for Business and Enterprise plans but broader rollout is expected. |
The common thread across all four: Codex is most powerful when the task is well-defined, the output is verifiable, and you do not want to stop what you are currently doing in order to do it. It is not a replacement for thinking. It is a way to execute while you think.