Codex Moved Into Claude Code, and That Says Everything

OpenAI released Codex as a Claude Code plugin on the same day Anthropic announced Computer Use. When I saw the repo, my first reaction was disbelief. A company shipping its agent inside a competitor’s agent. Effectively, anthropic-use.

Three lines to install it. Run /plugin marketplace add openai/codex-plugin-cc, then /plugin install codex@openai-codex, then /codex:setup. You need a ChatGPT subscription (free tier works) or an OpenAI API key, plus Node.js 18.18 or above. After that, you can call a second agent without leaving your Claude Code session.

What it actually does right now

This is not a full replacement for Claude Code’s built-in capabilities. Codex runs as a plugin layer on top of Claude’s default behavior, limited to two specific jobs: task delegation and code review.

/codex:review runs a read-only code review. /codex:adversarial-review stress-tests hidden assumptions. /codex:rescue hands control to Codex when Claude Code gets stuck. The --background flag lets Codex run asynchronously while you check progress with /codex:status.

The code review use case is where this gets interesting. Codex has been consistently rated above Claude for code review quality in community benchmarks. Plugging that strength directly into Claude Code’s workflow, where Claude handles the broader session, is a practical division of labor.

There is also a cost angle. If you connect a cheaper API instead of a Claude subscription to handle delegated tasks, token spend drops meaningfully. How much depends on your workload, but the architecture makes it possible.

Codex lost the taste battle

I have used both tools for months. The pattern I noticed was that Codex updates kept following Claude Code’s direction. Feature additions, UX patterns, even the sequencing of new capabilities. Claude Code’s team set the product taste, and Codex’s team trailed it.

Claude Code made this possible by design. The open plugin architecture meant anyone could extend it. That openness eventually pulled in a direct competitor. Codex could not beat Claude Code’s UX, so it stopped trying and moved inside instead.

The plugin ships with ChatGPT login integration, an app server model for connecting to any developer environment, and a fully open-source codebase. OpenAI is not being subtle about wanting Codex everywhere, even inside the rival’s house.

OpenAI’s real moat is not UX

After months of using Codex heavily, this move impressed me precisely because it acknowledged a weakness honestly. Claude Code’s developer experience is better. OpenAI conceded that point with a plugin instead of pretending otherwise.

What OpenAI read correctly is where its actual advantage sits. Claude’s reliability has real problems. Token policy changes have frustrated developers, and the error rates under load would be a liability issue in enterprise contexts. I keep using Claude anyway because the quality ceiling is high, but the floor is uncomfortably low.

OpenAI, meanwhile, ships frontier models at lower cost with higher volume and keeps error rates at production-grade levels. That infrastructure gap is something Chinese AI companies, despite impressive model quality, cannot replicate at the same scale.

Scaling keeps winning over scaffolding

Watching a company embed its product inside a competitor’s product can look like defeat. The “we lead this category” narrative weakens. But strip the optics away and this is the bitter lesson playing out in product strategy.

Codex entered the product it could not beat on UX and bypassed the taste gap entirely. Cursor, OpenCode, Droid, and every other scaffolding-first tool may lead on interface design. But the company that distributes frontier models cheaply and reliably at scale is the one that captures the ecosystem underneath all of them.

Sam Altman’s infrastructure-first bet pays off here. Regardless of which agent or IDE wrapper a developer chooses, the tokens consumed inside that tool have to come from somewhere. OpenAI is positioning itself as that somewhere, even when the surface-level product belongs to someone else.

The service layer is shifting ownership. The question for every developer tool company is whether they are building the house or supplying the electricity.