OpenAI Codex

Overview

Freshness note: AI products, model defaults, and usage limits change rapidly. This profile is a point-in-time snapshot last verified on July 10, 2026.

OpenAI Codex is the software-development agent in OpenAI’s product stack. On July 9, 2026, the standalone Codex app merged into the new ChatGPT desktop app for macOS and Windows. Codex keeps a dedicated mode for repositories, local files, terminals, diffs, pull-request review, and developer tools alongside Chat and ChatGPT Work.

Existing Codex app users can update normally and keep Codex as their default view and app icon. The CLI and IDE extension remain first-class developer surfaces, with shared config.toml settings for supported local configuration.

Key Features

Codex can inspect and edit multi-file repositories, run commands and tests, review diffs, and coordinate parallel subagents. The unified desktop release adds inline editing inside diffs, pull-request review in the side panel, faster GPT-5.6-powered Computer Use, and multi-repository projects.

The current model guide recommends the GPT-5.6 family for local Codex work. Power uses GPT-5.6 Sol with medium reasoning, Terra offers a balanced default, and Luna suits clear high-volume tasks. Max gives one agent more reasoning time, while Ultra uses subagents to split larger work in parallel.

Codex also spans the built-in browser, Computer Use, plugins, scheduled tasks, remote hosts, GitHub review workflows, and cloud execution. Those surfaces are related but not identical. Codex is not a selectable mode on ChatGPT web or mobile; supported desktop tasks can be reached from the Remote tab in the mobile app, while Codex cloud retains its own model and environment constraints.

Strengths

Codex is strongest when a task has a concrete objective, repository context, constraints, and a verifiable definition of done. It can handle the repetitive middle of engineering work—implementation, debugging, test scaffolding, review preparation, and documentation—while leaving architecture and release decisions with the human owner.

The unified app improves mixed workflows. A developer can keep code, browser verification, visual feedback, terminals, plugins, and review in one desktop task instead of moving state between unrelated assistants.

Limitations

Model availability differs by surface. The July model matrix exposes Sol, Terra, and Luna in the desktop app, CLI, and IDE extension, but not yet in Codex cloud; GPT-5.5 remains available there. Teams should test the exact client and authentication path instead of treating “Codex” as one universal runtime.

Agentic breadth also increases governance work. Computer Use, browser access, plugins, remote hosts, shell commands, and scheduled runs each have separate permission and approval boundaries. Codex can still produce plausible but wrong changes when a task is underspecified.

Practical Tips

Keep durable repository rules in AGENTS.md, then give each task a narrow objective, constraints, and acceptance criteria. Ask for a plan before large changes and review one implementation slice at a time.

For critical work, separate implementation and review passes. Inspect the diff, run the relevant tests, and keep pushes, deployments, external writes, and destructive actions behind explicit approval. Use Ultra only when the work has genuinely independent subproblems; parallelism is not a substitute for clean scope.

Verdict

OpenAI Codex remains one of the most complete coding-agent systems, now delivered as the dedicated development mode inside a broader ChatGPT desktop application. It is most valuable for teams that already practice clear specifications, disciplined review, and explicit permission boundaries.