Introducing Codex
OpenAI Codex: What is it?
- Codex is a software engineering agent that runs in the cloud and is made to handle several jobs at once.
- Codex-1 is a software engineering-focused version of OpenAI o3 that powers Codex.
- ChatGPT Pro, Team, and Enterprise users can access it now, and Plus users will soon be able to do the same.
How to use OpenAI Codex
- In ChatGPT, Codex is accessible via the sidebar.
- To assign tasks, type a prompt and select “Code.”
- You can click “Ask” for enquiries concerning your codebase.
- Every work is handled separately in a distinct, remote cloud sandbox environment.
- Your repository is already installed in this environment.
- Test harnesses, linters, and type checkers are among the commands that Codex can do in addition to reading and editing files.
- Depending on complexity, tasks usually take one to thirty minutes to complete. Progress can be tracked in real time.
- OpenAI Codex commits its modifications to its surroundings after a task is finished.
- It enables you to track every step by providing verifiable proof of its actions through citations of test outputs and terminal logs.
- After that, you have the option to examine the outcomes, submit more changes, launch a GitHub pull request, or incorporate the modifications straight into your local environment.
- You can set up the Codex environment to resemble your actual development environment.
The AGENTS.md files that are stored in your repository can serve as a guide for Codex. - Like README.md, these text files can provide Codex with information about project standard procedures, testing commands, and codebase navigation.
- Why OpenAI Codex agents function best with well-configured development environments, dependable testing configurations, and unambiguous documentation, much like human developers do. Nevertheless, codex-1 performs well even in the absence of custom scaffolding or AGENTS.md files.
Underlying Model and Training
Codex-1, an optimized version of OpenAI o3, powers Codex. It was trained using reinforcement learning on real-world coding tasks in a variety of environments with the aim of producing code that closely resembles human style and PR preferences, follows instructions exactly, and can iteratively run tests until passing results are obtained.
For use in Codex CLI, a scaled-down variant of codex-1 based on o4-mini is also accessible. This version is designed for low-latency code editing and Q&A, as well as quicker CLI processes.
Capabilities
The following are some of the tasks that OpenAI Codex can do:
- Write features
- Respond to enquiries about your codebase
- Fix bugs
- Submit pull requests for review
- Refactor
- Rename
- Write tests
- Scaffold new features
- Wire components
- Draft documentation
- Address on-call issues
- Plan tasks
- Offload background work
Safety and Security:
- In keeping with an iterative deployment approach, OpenAI Codex is made available as a research preview.
- Its design placed a high priority on security and transparency so that consumers could confirm its results.
- Citations, test results, and terminal logs allow users to verify Codex’s work.
- The agent expressly communicates any ambiguities or test failures when they occur.
- Users are cautioned that before integrating and running any agent-generated code, it must be carefully reviewed and validated.
- OpenAI Codex was taught to recognise and specifically reject requests meant to create malicious software. Even jobs that involve techniques occasionally employed for malware development (such as low-level kernel engineering) are distinguished and supported by it.
- Strict safety assessments and policy frameworks have been integrated.
- The agent runs exclusively in a safe, isolated cloud container for secure execution.
- Internet connection is blocked while a job is being executed, allowing only code supplied by GitHub repositories and user-configured dependencies to be used. External websites, APIs, and other services are not accessible via the agent.
Cost and Availability
- As of right now, Codex is being made available to ChatGPT Pro, Enterprise, and Team users worldwide.
- Plus and Edu users will shortly receive support.
- For the first few weeks, consumers will enjoy ample access at no extra charge.
- Rate-limited access and variable price alternatives for acquiring more usage will be implemented after this time.
- $1.50 per 1M input tokens and $6 per 1M output tokens, with a 75% prompt caching reduction, are the prices for developers using the codex-mini-latest model that is accessible through the Responses API.
Restrictions and the Future
- OpenAI Codex does not yet include capabilities like picture inputs for frontend work or the ability to course-correct the agent while it is operating, making it a research preview.
- Compared to interactive editing, which could need some tweaking, delegating to a remote agent can take longer.
- According to OpenAI, developers will assign tasks to agents in the future, enabling them to work more quickly and efficiently.
- It is anticipated that communicating with Codex agents would more closely resemble asynchronous teamwork.
- More flexible and interactive agent workflows will be introduced in the future, enabling users to cooperate on strategies, offer mid-task coaching, and get proactive updates.
- More thorough integrations between current tools are planned; in addition to GitHub, Codex CLI, ChatGPT Desktop, problem trackers, and continuous integration systems may be added in the future.
- The ultimate objective is for task delegation and real-time pairing to merge into a single process that works with all IDEs and commonplace tools.
Codex CLI
Codex CLI is an open-source, lightweight coding agent that operates within your terminal.
- It speeds up task completion by integrating the power of models like O3 and O4-mini into your local workflow.
- The default model in Codex CLI is a scaled-down version of codex-1, based on o4-mini and tailored for low-latency code Q&A and editing.
- It’s now simpler to link your developer account to the Codex CLI; you may log in using your ChatGPT account rather than creating API tokens by hand.
- Free API credits can also be redeemed by Plus and Pro customers who log into Codex CLI via ChatGPT.
In order to help engineers ship more quickly, OpenAI Codex seeks to eliminate context-switching, offload monotonous activities, and maintain concentration on important work. One of the first sectors to experience notable productivity increases from AI is software engineering.
