May 17 News.OpenAI today (May 17) announced the launch of cloud-based AI Agent CodexSpecialized in automationSoftware Development,Covering tasks such as fixing bugs and implementing features, officials hope the intelligence will usher in a new paradigm in code writing.
Codex is based on codex-1 modelOpenAI Codex is currently available as a research preview on the ChatGPT platform, with Pro, Enterprise and Team users being the first to experience it, and Plus and Edu users having to wait a while.
Developers can simply enter text commands into the ChatGPT sidebar to code in "Code" mode or ask code questions in "Ask" mode.
Each session of Codex runs in a separate cloud container, with the relevant codebase preloaded to ensure environmental isolation.AI intelligences can read and modify files, run tests, execute commands such as linters or type checkers, and log the results of their execution.
According to OpenAI, most tasks take between 1 and 30 minutes. Once the task is completed, Codex records the changes in detail through terminal logs and test results, making it easy for developers to verify each change. In addition, the development environment can be configured to match the team's actual development setup, and guidance is provided through the AGENTS.md file, which covers test specifications, code structure, and pull request (PR) information to ensure uniformity of rules within the project catalog.
Codex is already being used internally by OpenAI for tasks such as code refactoring, test generation, and bug fixing, and by early partners such as Cisco, Temporal, and Superhuman.
For example, Temporal uses Codex for error analysis and component connectivity; Superhuman lets product managers make small code changes on their own; and Kodiak uses Codex for test coverage improvement and debugging tool enhancements for self-driving software.
Internal benchmarks show that even without the AGENTS.md file or a customized environment, the codex-1 model performs well and gets the job done accurately.
Note: Codex does not currently support image input and is unable to interact with intelligences during task execution, the model has no internet access and is limited to the provided codebase and pre-installed dependencies.
Codex is currently free to use, and OpenAI expects to introduce flexible pricing in the future, with the codex-mini-latest model used by Codex CLI (Pairing of Lightweight Programming Intelligence in User Interfaces) already priced at $1.50 per million input tokens and $6 per million output tokens.