OpenAI Codex overview

OpenAI recently announced the launch of Codex, a cloud-based software engineering agent capable of handling multiple tasks simultaneously. Codex assists with writing features, answering questions about codebases, fixing bugs, and proposing pull requests for review, all within individual cloud sandbox environments preloaded with user repositories.

OpenAI Codex overview: AI-Powered Coding, Automation, and Developer Tools

Codex is powered by the codex-1 model, an optimized version of OpenAI’s o3 designed specifically for software engineering. It was trained using reinforcement learning on real-world coding tasks across various environments to generate code that aligns closely with human style and pull request preferences, strictly follows instructions, and iteratively runs tests until successful outcomes are achieved. Initial rollout targeted ChatGPT Pro, Enterprise, and Team users, with support for Plus and Edu tiers planned soon.

How Codex Works

Codex is accessible through ChatGPT’s sidebar, where users can assign new coding tasks by typing a prompt and selecting “Code,” or ask codebase-related questions with “Ask.” Each task runs in an isolated environment capable of reading, editing files, and executing commands like test harnesses, linters, and type checkers. Task durations range from one to thirty minutes, and users can monitor progress in real time.

Upon completion, Codex commits changes within its environment and provides verifiable evidence through citations, terminal logs, and test outputs, allowing users to trace every step. Users can review outcomes, request revisions, open GitHub pull requests, or integrate changes locally. Codex environments can also be configured to match real development setups.

Codex performance improves when guided by AGENTS.md files placed in repositories, offering instructions on navigation, testing commands, and adherence to project standards. Even without these files, codex-1 demonstrates strong performance on internal benchmarks.

Building safe and trustworthy agents

OpenAI emphasized its commitment to security and transparency by releasing Codex as a research preview. Codex includes safeguards such as verifiable outputs and explicit communication about uncertainties or test failures. Users are strongly encouraged to manually review and validate all agent-generated code before integration.

To prevent misuse, Codex was trained to recognize and reject requests related to malicious software development, while supporting legitimate low-level engineering tasks. Enhanced policy frameworks and rigorous safety evaluations have been introduced, as documented in an addendum to the o3 System Card.

Codex operates within a secure cloud container with no internet access during task execution, ensuring interaction is limited to provided repositories and pre-installed dependencies.

Early Use Cases and External Collaborations

OpenAI’s technical teams have incorporated Codex into daily workflows, using it for repetitive tasks like refactoring, renaming, test writing, feature scaffolding, bug fixing, and documentation drafting. Codex has also supported triaging on-call issues and background task management, helping engineers maintain focus and accelerate delivery.

Select external partners have been evaluating Codex across diverse environments:

• Cisco has explored using Codex to expedite ambitious engineering initiatives, providing feedback that helps shape future development.
• Temporal has used Codex to boost feature development, debug issues, run tests, and refactor large codebases, keeping engineers in flow.
• Superhuman has leveraged Codex to improve test coverage and fix integration issues, even allowing product managers to make lightweight code contributions.
• Kodiak has applied Codex to create debugging tools, enhance test coverage, and refactor code for its autonomous driving technology, while using it as a reference tool to understand complex stack components.
   

OpenAI recommended assigning well-defined tasks to multiple agents and experimenting with varied tasks and prompts to fully explore Codex’s capabilities.

Codex CLI Updates

In April 2025, OpenAI released Codex CLI, a lightweight open-source coding agent for terminal use, powered by models like o3 and o4-mini. On May 16, 2025, OpenAI introduced codex-mini-latest, a smaller o4-mini model optimized for low-latency code Q&A and editing in the CLI, now available as the default in Codex CLI and the API.

Additionally, developers can now link Codex CLI to their ChatGPT accounts, simplifying API key generation and configuration. Plus and Pro users can redeem $5 and $50 in free API credits, respectively, for the next 30 days.

Availability, Pricing, and Roadmap

Codex has been rolled out to ChatGPT Pro, Enterprise, and Team users globally, with upcoming access for Plus and Edu users. For several weeks, generous no-cost usage will be provided, followed by rate-limited access and flexible, on-demand pricing.

For developers, codex-mini-latest is priced at $1.50 per million input tokens and $6 per million output tokens, with a 75% discount for prompt caching.

As a research preview, Codex currently lacks features like image input for frontend development or real-time agent course correction, and its remote nature may introduce longer task durations compared to interactive editing. However, OpenAI envisions evolving Codex into an asynchronous collaborator capable of handling complex, long-running tasks.

Future Directions

OpenAI aims to build a suite of Codex tools supporting both real-time collaboration and asynchronous task delegation. The company anticipates a future where developers assign routine or complex tasks to agents, boosting productivity.

Upcoming features will include mid-task guidance, collaborative implementation strategies, proactive progress updates, and deeper integration with existing developer tools like GitHub, issue trackers, and CI systems.

As AI-driven productivity gains reshape software engineering, OpenAI is working with partners to understand the broader impacts on workflows, skill development, and global access. This is viewed as only the beginning of what Codex can achieve.