OpenAI’s next “low-key research preview” has arrived.
This time, it’s not ChatGPT, but a coding agent dubbed Codex that is being made available to ChatGPT Pro, Enterprise, and Team subscribers starting Friday. By drumming up comparisons to how ChatGPT was first described, CEO Sam Altman and other company leaders are positioning Codex as the company’s next major product. It doesn’t cost extra to use for now, though OpenAI plans to eventually charge for access once it gets a sense of demand.
The goal for Codex is to make ChatGPT a “virtual coworker” for engineers, Josh Tobin, OpenAI’s research lead for agents, said during a press call I attended this week. Like other vibe coding tools, Codex generates code from natural language. It can act independently on sandboxed code to fix bugs, run tests, and suggest changes to how code should run in the real world. This process can take anywhere up to 30 minutes, and OpenAI plans to let Codex work in the background for longer over tome.
Codex is integrated into ChatGPT’s web app to start, but it’s intentionally cut off from being able to access the internet to mitigate security risks. It’s powered by a version of OpenAI’s o3 reasoning model that is customized for coding and called codex-1.
According to Tobin, the company sees Codex as complementary to more granular AI coding assistants like Cursor and Windsurf, the latter of which OpenAI is in talks to acquire for roughly $3 billion. Inside OpenAI, Codex is already being used by engineers as a “morning to-do list” that helps them run multiple tasks in parallel. They have it spun up multiple tasks in parallel that they can come back to check on, according to Alexander Embiricos, Codex’s product lead. He said that a handful of companies that have tested it externally are seeing Codex used by on-call engineers who oversee a service’s stability.
For now, Codex is relatively limited in what it can do autonomously. Eventually, OpenAI’s goal is for it to fully abstract away the complexity of coding. “The way that we think most development will happen in the future is that the agent will work on its own computer, and we'll delegate to it,” said Embiricos. During a recent talk, Altman described coding as “central to the future of OpenAI.” There’s a belief in Silicon Valley that whoever creates a general-purpose AI engineer, which Codex is supposed to become, will have an edge in the race to build artificial general intelligence.
Codex was what OpenAI called its first AI coding tool way back in 2021, before ChatGPT was released. Now, models helping people code is perhaps the hottest area of AI, with Anthropic and others betting heavily on it as a business. On Thursday, Windsurf announced its own suite of coding models. And earlier this week, Google’s Gemini added the ability to connect to GitHub and announced AlphaEvolve, an AI coding agent specifically designed for developing algorithms.
Speaking of Google, the company’s annual conference, I/O, happens to be next week. Given the rivalry between OpenAI and Google, the timing of this week’s Codex announcement is likely not a coincidence. We’ll see how Google responds.
