GPT Codex: What the Name Actually Refers To in 2026

gpt codex is a query that lands on a moving target. This page sorts out what the name covers now, what the public numbers support, and what they do not.

Generate a project → Official site

The gpt codex name points at a product, not an endpoint

People type this query expecting something they can call. What they find is an agent: software that reads a repository, proposes changes, runs work and reports back, sold as part of a ChatGPT subscription rather than metered separately by the request. OpenAI dates its launch to April 2025 and describes the intervening period as a shift in how developers work, from asking a model for a snippet to delegating whole tasks and supervising several agents at once. That reframing matters for anyone budgeting, because it moves the unit of purchase from tokens to seats and allowances, and no amount of searching for a price per thousand tokens will produce a satisfying answer.

Reading 14.6 million installs honestly

The Visual Studio Code marketplace listing, published by OpenAI and free to install, reports about 14.6 million installs and 579 ratings averaging three stars out of five. Both halves deserve equal weight. An install count of that size says the tool reached a very large number of developers, which is genuine and hard to fake. A three-star average from a small sample of raters says the people motivated enough to leave a rating were split. Neither number tells you whether it suits your stack. The useful move is to read the most recent reviews in your own language and framework, treat the headline figures as context, and try it on a real task before forming a view.

The clearest product description is on a marketplace page

Vendor landing pages tend toward abstraction; the editor listing is unusually concrete. It describes adding the agent as a panel beside your code, where context from open files and selected text lets you write shorter prompts, and separately describes delegating bigger jobs to the cloud, tracking progress without leaving the editor, and opening cloud tasks locally for finishing touches. Sign-in works with a ChatGPT plan. On macOS the extension can also connect the desktop ChatGPT app to VS Code. If you want a plain statement of what the thing does before you commit to a trial, that listing and the documentation at developers.openai.com are the two places to read. Between them they answer what it does, where it runs and how you sign in, which is more than most product pages manage.

Access is gated by plan, not by key

Installing the extension is free; using the agent behind it is not. Public listings name Plus, Pro, Business, Edu and Enterprise as the plans it works with, and a February 2026 announcement added a limited-time inclusion with ChatGPT Free and Go alongside doubled rate limits on the paid tiers. The terminal client offers a second route, authenticating either with a ChatGPT plan or with an OpenAI API key, which is the closest thing to metered access available. Because these arrangements have already changed more than once, treat the current pricing pages as the only authority and treat any figure quoted in an article, including this one, as dated on arrival.

What none of these pages settle

Three questions survive a careful read of every official page. How much work a given plan actually affords, since limits are documented as varying by plan and by task rather than as a fixed cap. How the agent behaves on unfamiliar or unusual codebases, which no marketing page has ever answered for anyone. And whether the mixed ratings reflect early releases or current behaviour, which only recent reviews can indicate. Those gaps are not evasions so much as the genuinely hard parts, but they mean a trial on your own repository remains the only evaluation worth acting on. Budget an afternoon for that rather than another hour of reading, this page included.

Facts on the record

The questions people arrive with

How to evaluate it in an afternoon

1. Confirm your plan qualifies
   Check the current pricing page rather than an article. Plan inclusion and rate limits have changed at least once, including a limited-time extension to Free and Go accounts.
2. Install the free extension
   The VS Code listing is published by OpenAI and costs nothing, so the only commitment at this stage is the sign-in and the time it takes to point it at a project.
3. Give it a task you can grade
   Pick something with a clear right answer in code you know well: a failing test, a rename with edge cases, a small refactor. Vague tasks produce results nobody can score.
4. Read the diff, not the summary
   Judge the change it produced rather than its description of the change. That habit is what separates people who find agents useful from people who quietly stop using them.

Questions about the name and the product