Openai announced on Friday that it will begin a research preview of Codex, the company’s most capable AI coding agent.
Codex is powered by Codex-1, a version of the company’s O3 AI inference model optimized for software engineering tasks. According to Openai, Codex-1 generates “cleaner” code more than O3, glues the instructions more accurately, and runs the tests repeatedly on the code until you get the results.
The Codex agent runs on sandboxed virtual computers in the cloud. By connecting to GitHub, the Codex environment may be preloaded in the code repository. According to Openai, AI coding agents take 1-30 minutes to write simple features, fix bugs, answer questions about the codebase, run tests, and execute tasks.
Codex can handle multiple software engineering tasks simultaneously, Openai said, and does not restrict users from accessing their computers and browsers while they are running.
Codex is rolling out to ChatGpt Pro, Enterprise and Team subscribers starting today. Openai says users will begin “generous access” to Codex, but in the coming weeks, the company will implement rate limits for the tool. Users will then have the option to purchase additional credits to use Codex, an Openai spokesperson tells TechCrunch.
Openai plans to expand Codex Access to ChatGpt Plus and EDU users soon.
AI tools for software engineers, also known as Vibe Coders, have been gaining popularity over the past few months. CEOs of Google and Microsoft claim that around 30% of the company’s code is written by AI. In February, Anthropic released its own agent coding tool, Claude Code, and in April, Google updated its AI coding assistant Gemini Code Assist with more agent capabilities.
All of its atmosphere coding has made the business behind the AI coding platform part of the fastest growing technology. Cursor reportedly has raised new funds at a valuation of $9 billion, among its most popular AI coding tools, earning approximately $300 million per year in April.
Now Openai wants a part of the pie. The ChatGpt maker reportedly has signed a deal to acquire Windsurf, the developer behind another popular AI coding platform, for $3 billion. The launch of Codex is very clear that Openai is building its own AI coding tool.
Users with access to Codex can assign a new coding task for the agent by finding the tool in the ChatGPT sidebar, typing the prompt and clicking the Code button. Users can also ask questions about the codebase and click the (ASK) button. Under the prompt bar, users can see other tasks they have assigned codexes and monitor their progress.
At a briefing ahead of Codex’s launch, Openai’s agent research leader Josh Tobin told TechCrunch that the company hopes that the AI coding agent will ultimately act as a “virtual teammate” and that human engineers will autonomously complete tasks to achieve “hours or days.” Openai claims that it is already using Codex internally to offload repetitive tasks, new scaffolding features and draft documents.
Alexander Embiricos, a lead in Openai products, says many of the safety work of the company’s O3 model applies to codex. In a blog post, Openai says Codex will certainly reject requests to develop “malicious software.” Additionally, Codex works in an airy environment where there is no access to the wider Internet or external APIs. This limits Codex to being in the hands of a bad actor, but it can also hamper its usefulness.
It is worth noting that AI coding agents, like all generation AI systems today, are prone to mistakes. A recent Microsoft study found that industry-leading AI coding models such as the Claude 3.7 Sonnet and O3-Mini struggle to reliably debug software. However, it does not seem to undermine the excitement of investors in these tools.
Openai is also updating the Codex CLI, the company’s recently launched open source coding agent running in terminals using a version of the O4-MINI model optimized for software engineering. That model is currently the default for the Codex CLI, and will be available on Openai’s API for $1.50 per million input tokens (approximately 750,000 words, the entire Rod of the Rings Book series) and $6 per 1M output token.
The launch of Codex shows Openai’s latest efforts to enhance ChatGpt with additional products besides the infamous chatbot. Over the past year, Openai has added priority access as an advantage for the company’s AI video platform Sora, its research agents, deep search, and its web browsing agents, operators and subscribers.
These products could tempt more users to sign up for a ChatGPT subscription, and, in the case of Codex, specifically, persuade existing subscribers to pay more.