Today, we are introducing a research preview of Codex, a cloud-based software engineering agent that automates the creation of new features, provides direct answers about your codebase, identifies and fixes bugs, and generates pull requests for review. Each function operates in its isolated cloud sandbox, pre-configured with your repository.
Codex runs on Codex 1, a version of OpenAI O3 designed for software engineering. It was trained with reinforcement learning on real programming tasks in different environments. This helps it write code that fits human style and pull request preferences, follows instructions closely, and runs tests until they pass, ensuring early access and integration for users. We are making Codex available to ChatGPT Pro, Enterprise, and Business users today. Support for Plus and EDU users is coming soon.
How Codex Works
You can access Codex from the Chat GPT sidebar to start a coding task. Describe your goal and click Code. For questions about your code, use Ask. Each request launches in an isolated environment with your repository reloaded. Codex analyzes, edits files, and executes commands like test harnesses, linters, and type checkers. Task times range from one to thirty minutes, depending on complexity. Progress is visible in real time.
When Codex completes a task, it commits changes in its environment and displays terminal logs and test outputs as proof of every step. You can review results, request additional changes, open a GitHub pull request, or add the changes locally. You can also configure code. Products to closely mirror your real development environment.
Guide codecs with agents.md files in your repository. These files, like README.md, tell codecs how to work in your code, which testing commands to use, and how to follow your project’s practices. Codecs work best when developed by human developers in a well-defined environment, with reliable testing tools and clear documentation.
In coding tests and internal benchmarks, Codex 1 performs well even without agents.MD files or extra setup.
Building Safe and Trustworthy Agents
We’re releasing Codex as a research preview as part of our step-by-step deployment plan. We focused on security and transparency so users can check Codex’s output. This is especially important as Codex models take on more complex coding tasks. Users can review Codex’s work using citations, terminal logs, and test results if Codex is unsure or runs into test failures. It clearly lets you know so you can decide what to do next. It’s still important for users to review and check all code made by the agent before using it.
Aligning With Human Preferences
One of our main goals in training Codex 1 was to make HC outputs more closely match human coding preferences and standards than OpenAI. O3 Codex 1 creates cleaner patches that are ready for review and can be added to your workflow right away.
Preventing Abuse
It’s more important than ever to prevent harmful uses of AI in software engineering, like creating malware. At the same time, we need to ensure that security measures don’t block legitimate, useful work even when it uses similar techniques, such as low-level kernel engineering.
To keep codecs both secure and useful, we’ve trained it to spot and block harmful software requests while enabling productive, valuable work. Our updated policies and strict safety checks draw a clear line between the two. Details are in the O3 codex system card addendum.
Secure Execution
The Codex agent operates in a secure, isolated cloud environment. It works without internet access and uses only code from GitHub and your preset dependencies. It cannot reach external websites, APIs, or services.
Early Use Cases
OpenAI engineers rely on codecs daily to tackle repetitive tasks such as refactoring, renaming, and testing while staying focused on higher-value work. Codecs accelerate feature development, connect systems, resolve bugs, and draft documentation. Teams push productively further by sorting on-call issues, planning daily, and offloading background work. Codecs minimize context shifts and highlight neglected tasks, enabling engineers to move faster and focus on priorities.
- Cisco is testing Codex to accelerate project delivery. As early design partners, they shape Codex’s future by using it in real scenarios and providing direct feedback to OpenAI.
- Temporal uses codecs to speed up feature development, fix bugs, test, and refactor big code bases while also handling complex tasks. Disks in the background to keep engineers focused and moving faster.
- Superhuman deploys codecs for small, repetitive jobs like boosting test coverage and tackling integration issues, so product managers ship updates faster by making code changes, with codecs engineers stepping in only for code review.
- Kodiak leverages codecs to build debugging tools, raise test coverage, and optimize code. Advancing its Kodiak driver for autonomous driving, Codex serves as a crucial reference, providing context and version history to enhance the engineers’ understanding.
Based on current insights, assign clear, well-defined tasks to multiple agents for maximum efficiency. Experiment with diverse tasks and prompts to fully explore the model’s capabilities.
Updates To Codex CLI
We recently launched Codex CLI, an open-source coding agent for your terminal. It brings models like O3 and O4 Mini directly into your workflow, speeding up task completion.
We are now launching a slimmer Codex 1 based on the O4 Mini, custom-built for Codex CLI. This model accelerates CLI workflows, delivering fast Q&A, quick edits, and strong adherence to instructions and style. It’s the CLI default and available in the API as Codex Mini, latest with ongoing updates.
Connecting your developer account to Codex CLI is now seamless. Simply sign in with your ChatGPT account and choose your API organization. We handle the API key setup, and Plus and Pro users signing in unlock $5 and $50 in free API credits over the next 30 days.
Codex Availability Pricing And Limitations
Starting today, Codex is available globally to ChatGPT Pro, Enterprise, and Business User Plus, and EDU access will follow soon. For the next few weeks, Codex is free to try; afterward, we’ll introduce rate limits and flexible pricing buy extra usage as you need.
Building with Codecs mini: latest. Access it via the Responses API. Pricing is $1.50 per 1K input tokens and $6 per 1K output tokens, with 75% off for prompt caching.
Codex is still in the early stages right now as a research preview. It doesn’t support features like image inputs for front-end work or mid-task adjustments to the agent’s work. Also, sending tasks to a remote agent takes longer than editing directly. So it may take some time to get used to it. In the future, working with Codex agents will feel more like collaborating with colleagues asynchronously. As the model improves, we expect agents to handle even more complex tasks for longer periods.
What’s Next?
G Envision developers focused on what matters with agents handling the rest to maximize productivity to achieve this. We are building codec tools for seamless real-time collaboration and asynchronous task delegation.
AI tools like Codex CLI now set the industry pace, helping developers code faster. The new asynchronous multi-agent Codex workflow for ChatGPT is now on track to become the industry standard for high-quality engineering.
Ultimately, real-time pairing and task delegation will converge. Developers and AI agents will seamlessly collaborate on ideas, tools, and daily tasks asking questions, receiving suggestions, and delegating complex work in a unified flow.
Looking forward, we plan to add more interactive and flexible ways to work with agents. Soon, developers will be able to provide feedback during tasks, collaborate on strategies, and receive updates as work progresses. We also want to connect Codex with more of your favorite tools. Right now, Codex works with GitHub, and soon you will be able to assign tasks from Codex, the CLI, ChatGPT, the Desktop, and even your issue tracker in our CI system.
AI is driving rapid productivity gains across software engineering, unlocking new possibilities for individuals and small teams. While we are eager for what’s ahead, we collaborate with partners to deepen our understanding of agencies’ impact on workflows and skill development at every level.
This is just the start, and we are excited to see what you create with Codex.
Source: ntroducing Codex
