OpenAI Launches Codex, Its Cloud-Based Coding Agent

An Agent That Doesn't Autocomplete—It Works Autonomously

OpenAI today introduced Codex, a cloud-based software engineering agent that can work on multiple coding tasks in parallel. The company describes it as a "research preview," available starting today to ChatGPT Pro, Business, and Enterprise users, with support for Plus and Edu users coming soon, according to OpenAI's announcement.

The leap beyond conventional code-autocomplete tools is significant. Codex doesn't suggest the next line as a developer types—it takes on a complete task (writing a feature, fixing a bug, answering questions about a repository) and carries it out from start to finish in its own environment, without continuous supervision.

How It Works in Practice

According to OpenAI, each task assigned to Codex runs in its own isolated cloud "sandbox," preloaded with the user's repository. From the ChatGPT sidebar, users simply type a prompt and click "Code" to set the agent to work, or "Ask" if they just want to ask a question about existing code.

Inside that environment, Codex can read and edit files and run commands such as test suites, linters (tools that check code style and flag potential errors), and type checkers. OpenAI notes that completing a task typically takes between 1 and 30 minutes, depending on its complexity, and that progress can be monitored in real time.

Once the work is done, Codex commits its changes in its own environment and provides what OpenAI calls "verifiable evidence": citations of terminal logs and test outputs that let users trace every step taken during the task. From there, developers can review the results, request further revisions, open a GitHub pull request directly, or integrate the code into their local environment.

OpenAI has also introduced AGENTS.md files, a format similar to README files that tells Codex how to navigate a given repository, which commands to use for testing, and which conventions the project follows. The company says that even without these guide files or any custom scaffolding, codex-1 delivers strong performance on its internal coding evaluations.

The Model Behind It: codex-1

Codex runs on codex-1, a version of OpenAI o3 optimized specifically for software engineering. According to OpenAI, the model was trained using reinforcement learning on real-world coding tasks across a variety of environments, with the goal of generating code that closely mirrors human style and PR preferences, follows instructions precisely, and can iteratively run tests until it gets a passing result.

A primary goal while training codex-1, the company explains, was to align the model's outputs closely with human coding preferences and standards. Compared to OpenAI o3, codex-1 consistently produces cleaner patches, ready for immediate human review and integration into standard engineering workflows.

Security and Sandboxed Execution

Part of the announcement is devoted to explaining how OpenAI has addressed the risks of an agent capable of writing and running code autonomously. The agent operates entirely within a secure, isolated cloud container. During task execution, internet access is disabled, so the agent's interaction is limited to the code explicitly provided through GitHub repositories and to dependencies preinstalled via a user-defined setup script. The agent cannot access external websites, APIs, or other services.

Regarding the risk of malicious use, such as malware development, OpenAI says Codex was trained to accurately identify and reject such requests, distinguishing them from legitimate tasks that sometimes rely on similar techniques, such as low-level kernel engineering. The company has strengthened its policy frameworks and added extra safety evaluations, and it has published an addendum to o3's "system card" to reflect these assessments.

Even with these safeguards, OpenAI stresses one point: it still remains essential for users to manually review and validate all agent-generated code before integrating or running it.

Who's Already Using It

According to OpenAI, its own engineering teams have already started using Codex in their day-to-day work, mainly for repetitive, well-defined tasks—refactoring code, renaming elements, writing tests—that would otherwise break their concentration. The company also says it's proving useful for scaffolding new features, wiring up components, fixing bugs, and writing documentation.

Ahead of launch, OpenAI worked with a small group of outside testers to understand how Codex performs across different codebases and development workflows. Among the cases the company cites, Cisco is evaluating the agent as an early design partner for product ideas across different parts of its portfolio. Temporal uses it to speed up feature development, debug issues, write and run tests, and refactor large codebases, while keeping its engineers focused by running complex tasks in the background. Superhuman uses it to speed through small but repetitive tasks, such as improving test coverage or fixing integration bugs, letting product managers make light code changes without needing an engineer, except for final review. Kodiak, for its part, uses Codex to write debugging tools, improve test coverage, and refactor code as it develops its self-driving technology, and also as a reference tool to help its engineers understand parts of the system they're less familiar with.

Drawing on the experience of these early testers, OpenAI recommends assigning well-defined tasks to multiple agents at once and experimenting with different task types and prompt phrasings to get a better sense of the model's capabilities.

A Lightweight Version for the Terminal

Alongside the Codex launch, OpenAI notes that it introduced Codex CLI last month, a lightweight, open-source coding agent that runs directly in a user's terminal, bringing models like o3 and o4-mini into the local workflow. Today the company is also announcing a smaller version of codex-1, based on o4-mini and built specifically for Codex CLI, optimized for low-latency code questions and edits within that environment.

What Changes for Developers

The difference from a traditional autocomplete assistant lies in the degree of delegation involved. This isn't about accepting or rejecting a suggestion line by line—it's about handing off an entire task and getting back, minutes later, a tested set of changes documented with its own execution logs. That shifts the developer's job toward defining the task and reviewing the outcome, rather than writing the code itself.

OpenAI is framing the launch as a research preview rather than a finished product, and for now access remains limited to certain subscription tiers. The company makes clear that human validation of the generated code remains a mandatory step—not an optional one—before shipping it to production.