OpenAI ships GPT-5-Codex

OpenAI released GPT-5-Codex, a version of GPT-5 trained specifically for agentic coding, and made it the default model for Codex’s cloud tasks and code review, with access also available through the Codex CLI and IDE extension. Rather than tuning for single-turn completions, OpenAI trained the model on real-world engineering work — debugging, writing tests, building projects from scratch and executing large-scale refactors — with the aim of having it act more like an autonomous collaborator than a completion tool.

OpenAI and outside reporting both described the model as able to work independently on a task for extended periods, running code, checking its own output and adjusting course without further prompting, for up to several hours at a stretch. A separate capability, code review, had the model navigate a codebase, reason about dependencies between files and run tests to check whether a proposed change actually worked, rather than reviewing a diff in isolation. OpenAI said it had already put this reviewer to work on its own internal pull requests, where it reported catching a large number of issues daily.

The release extended a pattern already visible with GPT-5: rather than one general-purpose model, OpenAI shipped a coding-specialised variant alongside it, betting that agentic software engineering — models that plan, act and check their own work over long horizons — was distinct enough from conversational use to warrant separate training and separate product surfaces.