It runs autonomously in the cloud and can connect to the user’s GitHub account.
Codex at work. Source: OpenAI
- The AI agent runs on a new model called codex-1, a version of the o3 reasoner optimized for development. It writes “cleaner” code and tests it until it works, OpenAI says.
- codex-1 has a context length of 192 thousand tokens and a “medium reasoning level”.
- The Codex agent runs in the cloud, “without occupying” the user’s browser, and can be launched from a smartphone. The AI assistant can perform several tasks simultaneously and also connect to GitHub repositories. Once completed, it displays a log of all changes to the code.
- OpenAI has made the AI Agent available to Pro, Enterprise, and Team subscribers. The Codex will appear in the sidebar in ChatGPT and will open in a separate window.
- The company said there would be no usage restrictions at first, but that it would introduce speed limits for Codex “in the coming weeks.” After that, a system for purchasing additional credits would be introduced. Access for Plus and Edu plan users is planned for later.
- The developers also released a codex-mini model based on o4-mini. It will be available in the API at a price of $1.5 and $6 for 1 million input and output tokens.