OpenAI Codex in ChatGPT: AI Coding Agent Launched

OpenAI Introduces Codex: A New AI Coding Agent

On Friday, OpenAI announced the research preview release of Codex, representing their most advanced AI solution for coding tasks to date.

Codex's Underlying Technology

Codex is built upon codex-1, a specialized iteration of OpenAI’s o3 AI reasoning model. This version has been specifically optimized for the complexities of software engineering. OpenAI asserts that codex-1 generates more refined code compared to o3.

Furthermore, it demonstrates greater adherence to provided instructions and employs iterative testing procedures to ensure successful code execution.

Operational Environment and Capabilities

The Codex agent functions within a secure, virtualized computing environment hosted in the cloud. It can integrate with GitHub, allowing pre-loading of existing code repositories for seamless access.

OpenAI estimates that this AI coding agent can complete tasks – from writing simple features to debugging, answering codebase questions, and running tests – within a timeframe of one to thirty minutes.

Codex is capable of managing multiple software engineering tasks concurrently, without restricting user access to their computer or web browser during operation.

Availability and Access

The rollout of Codex begins today for subscribers to ChatGPT Pro, Enterprise, and Team plans. Initially, users will receive “generous access” to the tool.

However, OpenAI intends to implement rate limits in the coming weeks. Subscribers will then have the option to purchase additional credits to extend their usage of Codex, as confirmed by an OpenAI spokesperson to TechCrunch.

Plans are underway to broaden access to Codex to include ChatGPT Plus and Edu users in the near future.

The Rise of AI Coding Tools

AI-powered tools designed for software engineers, often referred to as “vibe coders,” have experienced a significant surge in popularity recently.

Executives at both Google and Microsoft have reported that approximately 30% of their companies’ code is now generated by AI. Anthropic launched Claude Code in February, and Google enhanced its Gemini Code Assist with more agentic features in April.

Market Growth and OpenAI's Strategy

The increasing demand for these AI coding platforms has fueled rapid growth within the businesses that develop them. Cursor, a leading AI coding tool, achieved an annualized revenue of around $300 million in April and is currently seeking new funding at a $9 billion valuation.

OpenAI is now aiming to capture a share of this expanding market. Reports indicate that OpenAI has finalized an agreement to acquire Windsurf, the creator of another popular AI coding platform, for $3 billion.

The launch of Codex further demonstrates OpenAI’s commitment to developing its own suite of AI coding tools.

Using Codex within ChatGPT

Users who have access to Codex can locate the tool within the ChatGPT sidebar. New coding tasks can be assigned by entering a prompt and selecting the “Code” button.

Questions regarding the codebase can be posed by utilizing the “Ask” button. A display below the prompting bar allows users to monitor the progress of assigned tasks.

Future Vision and Safety Measures

Josh Tobin, OpenAI’s Agents Research Lead, shared with TechCrunch that the company envisions its AI coding agents evolving into “virtual teammates.” These agents would autonomously handle tasks that currently require hours or even days for human engineers to complete.

OpenAI is already leveraging Codex internally to automate repetitive tasks, create feature scaffolding, and generate documentation.

Alexander Embiricos, OpenAI Product Lead, emphasized that the safety protocols developed for the o3 model are also applicable to Codex. OpenAI states that Codex will consistently decline requests to create “malicious software.”

The agent operates within a secure, isolated environment, lacking access to the public internet or external APIs. This limitation, while enhancing security, may also affect its overall utility.

Limitations and Ongoing Development

It’s important to acknowledge that, like all current generative AI systems, AI coding agents are susceptible to errors. A recent Microsoft study revealed that leading AI coding models, including Claude 3.7 Sonnet and o3-mini, encountered difficulties with reliable software debugging.

Despite this, investor enthusiasm for these tools remains strong.

Updates to Codex CLI and API

OpenAI is also updating Codex CLI, its recently released open-source coding agent for terminal use, with a version of its o4-mini model. This model is now the default within Codex CLI and will be available through OpenAI’s API at a cost of $1.50 per 1M input tokens and $6 per 1M output tokens.

Expanding the ChatGPT Ecosystem

The launch of Codex represents OpenAI’s continued effort to expand ChatGPT’s functionality beyond its core chatbot capabilities. Recent additions include priority access to Sora (AI video platform), Deep Research (research agent), and Operator (web browsing agent) as benefits for subscribers.

These offerings are designed to attract new users to ChatGPT subscriptions and incentivize existing subscribers to purchase increased rate limits, particularly for tools like Codex.