OpenAI unveils GPT-5-Codex with GitHub reviews and IDE tools

OpenAI has released GPT-5-Codex, a variant of its GPT-5 family optimised for use with the Codex coding agent. The release coincides with updates across the Codex product stack, including a rebuilt Codex CLI focused on agentic workflows, an IDE extension for VS Code and compatible forks, GitHub integration for code reviews, and web workflow improvements.

Access to Codex is included for OpenAI users on ChatGPT Plus, Pro, Business, Edu, and Enterprise plans, with usage limits that scale by plan. OpenAI says Plus, Edu, and Business plans will cover “a few focused coding sessions each week,” while users on the Pro plan can expect a “full workweek across multiple projects.” API access is planned to arrive soon for Codex users who prefer programmatic billing.

GPT-5-Codex: model design and persistence

GPT-5-Codex is configured with a router and is presented as a self-contained model for agentic software engineering. OpenAI states the model uses a dynamic amount of reasoning and, in internal tests, “the model was able to work autonomously on a problem for more than seven hours (though that’s not a hard cap).” The release also formalises AGENT.md as the standard file for supplying coding guidelines and instructions to the agent.

GPT-5-Codex: developer workflows and tooling

The Codex CLI and IDE extensions aim to support agentic workflows and developer pairing. Developers can attach screenshots, wireframes, and diagrams to provide context, and the CLI gains a to-do list and clearer tool-call UI for following the agent’s actions and diffs. Approval modes are explicit: read-only with explicit approval, auto with full workspace access but requiring approvals outside that workspace, and full access with Code,x able to read files anywhere and run commands with network access.

OpenAI has introduced a GitHub code-review agent powered by GPT-5-Codex. “GPT-5-Codex has been trained specifically for conducting code reviews and finding critical flaws,” OpenAI explained. The company reports that GPT-5-Codex produced fewer incorrect comments and a higher share of high-impact comments in its internal evaluations: while 13.7% of comments left by GPT-5 (high) were incorrect, only 4.4% of GPT-5-Codex comments were wrong; 52% of GPT-5-Codex comments were judged “high-impact” versus 39% for GPT-5 (high). OpenAI adds that the agent can be invoked in pull requests to perform targeted reviews.

OpenAI reports a 74.5% score for GPT-5-Codex on the SWE-bench set referenced in its release, highlighting stronger performance on refactoring tasks compared to GPT-5 in high reasoning mode. Platform changes to the Codex web include caching improvements that the team says reduce median completion time for new tasks and follow-ups by 90%.

OpenAI positions GPT-5-Codex as a tool for both interactive pairing and extended autonomous work. “The model combines two essential skills for a coding agent: pairing with developers in interactive sessions, and persistent, independent execution on longer tasks,” the company writes in its announcement. “That means Codex will feel snappier on small, well-defined requests or while you are chatting with it, and will work for longer on complex tasks like big refactors.”

For teams considering adoption, key considerations will include plan-based usage limits, the choice of CLI approval mode, and how to integrate Codex review workflows into existing CI/CD and code-quality processes.

As OpenAI notes of its internal use, “At OpenAI, Codex now reviews the vast majority of our PRs, catching hundreds of issues every day — often before a human review begins. It’s been key to letting the Codex team move fast with greater confidence,” the company says.