OpenAI has announced the release of Codex, a cloud-based software engineering agent designed to assist developers by performing parallel coding tasks such as writing features, fixing bugs, answering questions about codebases, and proposing pull requests. The tool is now available as a research preview to ChatGPT Pro, Enterprise, and Team users, with Plus and Edu support expected soon.
Codex is powered by codex-1, a version of OpenAI’s o3 model optimized specifically for software development. According to OpenAI, codex-1 was trained using reinforcement learning on real-world coding scenarios to closely match human programming styles, follow instructions precisely, and test outputs until they pass.
How Codex Works
Users can access Codex via the ChatGPT sidebar by entering prompts or code questions and assigning tasks. Each task is executed in a secure, isolated cloud sandbox environment preloaded with the user’s repository. Codex can read, edit, and test code using tools like test harnesses, linters, and type checkers. Task execution time ranges from 1 to 30 minutes, and users can monitor progress in real time.
Once a task is complete, Codex commits its changes within its environment and provides verifiable output through terminal logs and test results. These outputs can be reviewed, revised, or integrated into GitHub pull requests or local codebases. Developers can configure Codex to closely mirror their actual dev environments.
Codex Features and Developer Controls
Developers can guide Codex’s behavior using AGENTS.md files—similar to README.md—to define project-specific instructions, command preferences, and testing procedures. While Codex performs best with customized environments and documentation, internal testing shows it maintains strong performance even without tailored scaffolding.
Codex tasks operate independently and securely, with no internet access during execution. This isolation ensures safety and containment of each task.
To address safety concerns, OpenAI says Codex was trained to distinguish between legitimate and malicious tasks and is equipped to explicitly refuse requests associated with malware development. Users can verify agent outputs via logs, citations, and test results, and Codex alerts users when it encounters test failures or uncertainties. However, OpenAI emphasizes the continued need for manual review before deploying agent-generated code.
Codex CLI and Lighter Model Variant
Alongside the core Codex launch, OpenAI is updating the Codex CLI, a local terminal agent tool. A smaller, faster version of codex-1—called codex-mini-latest—is now available. This version is optimized for low-latency Q&A and code editing and can be accessed via the Responses API, priced at $1.50 per 1M input tokens and $6 per 1M output tokens, with a 75% discount for prompt caching.
Codex CLI users can now link their developer accounts using ChatGPT sign-in, simplifying API setup. Starting today, Plus and Pro users can redeem $5 and $50 in free API credits, respectively, for the next 30 days.
Early Use Cases and Limitations
Internally, OpenAI teams are already using Codex to handle background tasks such as refactoring, documentation, debugging, and writing tests, reducing context-switching and improving team focus. While Codex has proven useful for a variety of real-world tasks, OpenAI acknowledges its limitations. The agent currently lacks image input support, real-time correction during execution, and may take longer to complete tasks compared to interactive editing.
As OpenAI continues to evolve Codex, the company anticipates a shift toward asynchronous collaboration between human developers and autonomous agents capable of managing longer and more complex software engineering responsibilities.
Codex is being released in line with OpenAI’s iterative deployment strategy and will be free to eligible users for a limited time before transitioning to rate-limited access and flexible pricing options.
Image: OpenAI
本文， ”OpenAI Launches Codex” 首次发表于 小型企业趋势