OpenAI launches Codex, a cloud-based coding agent
Built on a fine-tuned o3, each task runs in its own preloaded cloud sandbox and proposes a pull request, letting several jobs run at once without a developer at the keyboard.
- Models & capabilities
- Major
OpenAI launched a research preview of a new product called Codex, reusing the name of the code-generation model it had published in 2021 (and which underpinned GitHub Copilot) for a different kind of tool: a cloud-based software engineering agent rather than an autocomplete model. Built on a version of OpenAI’s o3 reasoning model fine-tuned for coding, the new Codex is given access to a copy of a developer’s repository inside an isolated cloud sandbox and asked to complete a task — fix a bug, write a feature, answer a question about the codebase — largely unsupervised, proposing a pull request for review when it finishes.
Because each task runs in its own sandbox, OpenAI said users could dispatch several jobs to Codex at once and let them run in parallel, rather than working through requests one at a time as with a chat-based coding assistant. The product shipped initially to ChatGPT Pro, Team and Enterprise subscribers, with Plus and edu access following within weeks.
The relaunch illustrated how far “coding assistant” products had moved since 2021: where the original Codex model produced code suggestions inline as a developer typed, the new product operated as a semi-autonomous agent that could plan, execute, and test changes without continuous human direction. It arrived alongside comparable agentic coding products from other labs and startups, intensifying competition over which company’s model could most reliably complete real software-engineering tasks rather than merely suggest snippets.