OpenAI has launched a research preview of Codex, its most advanced AI coding agent to date.
Codex is built on codex-1, a customized variant of the company’s o3 model, fine-tuned specifically for software development tasks. According to OpenAI, codex-1 generates cleaner code, follows instructions more accurately, and automatically runs tests until successful results are achieved.
The Codex agent operates within a cloud-based, sandboxed virtual environment. By connecting to GitHub, it can preload your code repositories, making setup seamless. OpenAI says Codex can perform a range of tasks—such as building simple features, fixing bugs, answering code-related questions, and executing tests—in anywhere from one to thirty minutes.
OpenAI notes that Codex can manage multiple development tasks at once and doesn’t interfere with your ability to use your computer or browser while it works.
Starting today, Codex is available to ChatGPT Pro, Enterprise, and Team subscribers. OpenAI says initial access will be “generous,” though rate limits will be introduced in the coming weeks. After that, users can purchase additional credits to continue using the tool.
Access for ChatGPT Plus and Edu users is expected to follow soon.
AI-powered tools for developers—often referred to as “vibe coders”—have seen explosive growth. Tech leaders at Google and Microsoft estimate that around 30% of their companies’ code is now AI-generated. Other players have entered the space too, including Anthropic with Claude Code and Google with the updated Gemini Code Assist.
These tools have fueled rapid growth for AI coding platforms. Cursor, for example, reportedly reached $300 million in annualized revenue by April and is seeking funding at a $9 billion valuation.
Now OpenAI is positioning itself more firmly in this arena. It recently closed a $3 billion acquisition of Windsurf, the team behind another popular coding platform. The release of Codex is a clear signal that OpenAI intends to build out its own suite of developer tools.
Users with Codex access will find it in ChatGPT’s sidebar. To assign coding tasks, they can type a prompt and click the “Code” button. For questions about their codebase, the “Ask” button can be used. Task progress is displayed below the prompt bar for easy tracking.
Ahead of the launch, OpenAI’s Agents Research Lead, Josh Tobin, said the long-term goal is for Codex to act as a “virtual teammate,” capable of completing tasks that would normally take engineers hours or even days. OpenAI has already begun using Codex internally to automate repetitive work, build new features, and draft documentation.
Product Lead Alexander Embiricos emphasized that much of the safety work done for the o3 model also applies to Codex. According to OpenAI, Codex is designed to refuse any requests to create harmful software. It operates in an isolated, air-gapped environment with no access to the internet or external APIs, reducing the risk of misuse—though this may also limit certain capabilities.
Like other generative AI tools, Codex isn’t flawless. A recent Microsoft study showed that top-tier models like Claude 3.7 Sonnet and o3-mini still struggle with debugging. Nevertheless, investor interest in AI coding agents remains strong.
Additionally, OpenAI is updating Codex CLI—its open-source terminal-based coding assistant—with a version of the o4-mini model optimized for development tasks. This model is now the default for Codex CLI and will be available via API at a cost of $1.50 per 1 million input tokens and $6 per 1 million output tokens.
The debut of Codex highlights OpenAI’s continued push to expand ChatGPT beyond its core chatbot functionality. Over the past year, it has added features like priority access to Sora (AI video), Deep Research, and the Operator browsing agent—all perks for subscribers.
These additions aim to draw more users into ChatGPT’s paid tiers—and in Codex’s case, possibly encourage existing users to spend more for higher usage limits.
