On Friday, OpenAI announced the launch of a research version of Codex, the company’s most powerful AI coder.
Codex is based on codex-1, a version of the company’s o3 AI reasoning model optimized for software engineering tasks. OpenAI claims that codex-1 produces “cleaner” code than o3, follows instructions more accurately, and will iteratively run tests on its code until a passing result is achieved.
The Codex agent runs in a virtual computer in a sandbox in the cloud. By connecting to GitHub, the Codex environment can be preloaded with your code repositories. OpenAI claims that the AI coding agent will take between one and 30 minutes to write simple functions, fix bugs, answer questions about your codebase, and run tests, among other tasks.
Codex can perform multiple software development tasks simultaneously, OpenAI says, and it does not restrict users from accessing a computer and browser while it is running.
Codex is launching today for ChatGPT Pro, Enterprise, and Team subscribers. OpenAI says that users will initially have “generous access” to Codex, but in the coming weeks, the company will introduce limits on the tool’s use. Users will be able to purchase additional credits to use Codex, an OpenAI spokesperson told TechCrunch.
Soon, OpenAI plans to expand access to Codex for ChatGPT Plus and Edu users.
Artificial intelligence tools for software engineers, also known as vibro coders, have increased dramatically in popularity in recent months. Google and Microsoft executives claim that approximately 30% of their companies’ code is now written with the help of AI. In February, Anthropic released its own agent-based coding tool Claude Code, and in April, Google updated its Gemini Code Assist AI coding tool to include more agent-based capabilities.
All of this has made the companies behind AI coding platforms some of the fastest growing in the tech industry. Cursor, one of the most popular AI coding tools, reached annual revenues of about $300 million in April and is reportedly raising new funds at an estimated $9 billion.
Now, OpenAI wants a piece of the pie. The maker of ChatGPT has reportedly entered into a deal to acquire Windsurf, the developer of another popular AI coding platform, for $3 billion. The launch of Codex clearly shows that OpenAI is also developing its own AI coding tools.
Users with access to Codex can find the tool in the ChatGPT sidebar and assign new coding tasks to the agent by typing in a query and clicking the “Code” button. Users can also ask questions about their codebase by clicking the “Ask” button. Below the prompts bar, users can see other tasks assigned to Codex and track their progress.
At a briefing ahead of Codex’s launch, Josh Tobin, OpenAI’s head of agent research, told TechCrunch that the company wants its AI coding agents to act as “virtual teammates,” autonomously completing tasks that would take “hours or even days” for human engineers. OpenAI says it already uses Codex internally to offload repetitive tasks, build new features, and prepare documentation.
OpenAI’s head of product, Alexander Embirikos, says a lot of the security work for the company’s o3 model also applies to Codex. In a blog post, OpenAI claims that Codex will reliably reject requests to develop “malicious software.” Additionally, Codex operates in an air-gapped environment, with no access to the internet or external APIs. This limits the danger Codex could pose in the hands of attackers, but could also hurt its usefulness.
It should be noted that AI coding agents, like all modern generative AI systems, are prone to errors. A recent study by Microsoft found that leading AI coding models like Claude 3.7 Sonnet and o3-mini cannot reliably debug software. However, this doesn’t seem to be deterring investor interest in these tools.
OpenAI is also updating Codex CLI, a recently launched open-source coding agent that runs in your terminal, with a version of its o4-mini model optimized for software engineering. This model is now the default in the Codex CLI and will be available in the OpenAI API for $1.50 per 1 million input tokens (roughly 750,000 words, more than the entire Lord of the Rings series) and $6 per 1 million output tokens.
The launch of Codex marks OpenAI’s latest attempt to beef up ChatGPT with additional products beyond its infamous chatbot. Last year, OpenAI added priority access to the Sora AI video platform, the Deep Research research agent, and the Operator web browsing agent as subscriber benefits.
These offers could attract more users to subscribe to ChatGPT and, in the case of Codex, convince existing subscribers to pay OpenAI more money for increased pricing caps.
