OpenAI Codex

Our take

Codex is OpenAI's answer to a genuinely different question than most AI coding tools are asking. Where Cursor and Claude Code keep you close to the action, Codex is built around stepping back. You describe a task, hand it off to a cloud agent, and come back when there's a pull request waiting for you. It's less about AI-assisted editing and more about AI-delegated engineering.

The architecture makes this possible and constraining in equal measure. Each Codex task runs in an isolated cloud sandbox with your GitHub repo cloned into it, network access disabled by default, and its own execution environment. The agent writes code, runs your test suite, and opens a PR. You review the diff, run CI, and merge or iterate. If the task is well-specified and the codebase is greenfield-friendly, this genuinely works.

The model quality is high. Codex runs on codex-1, a version of o3 tuned specifically for software engineering, and it shows on structured tasks: feature additions, bug fixes with clear reproduction steps, and refactors with a narrow scope. On SWE-bench style evaluations, it scores competitively with Claude Code. The gap between the two tends to open up on tasks that require nuanced reasoning across unusual codebases, or where the developer wants to stay interactively involved.

The fire-and-forget model is a genuine advantage for the right kind of work, and a source of friction for the wrong kind. If you need to run three agents in parallel on three separate features while you focus elsewhere, Codex handles that elegantly. If you want to watch what the agent is doing, steer it in real time, or work from local state that isn't committed to GitHub yet, you'll find the cloud-first architecture gets in the way.

The CLI is worth knowing about separately. OpenAI open-sourced it in April 2025 and it's built in Rust, so it's fast and lightweight. It handles local, interactive terminal sessions the same way Claude Code does. MCP is supported. The CLI is genuinely solid, but it's younger than Claude Code's and the surrounding ecosystem (hooks, skills, extended tooling) hasn't accumulated as much depth yet.

Pricing is tied to ChatGPT subscriptions, which is natural if you already use ChatGPT Pro or Business, and somewhat awkward if you don't. There's no standalone Codex plan. The API path currently centers on codex-mini-latest pricing rather than a standalone Codex subscription: $1.50 per million input tokens, $6 per million output tokens, and a 75% prompt-caching discount. That makes automation economics cleaner than the ChatGPT bundle, but the broader Codex product is still a cloud-agent workflow first.

For teams running a GitHub-centric workflow who want to try delegating whole tasks to an AI agent, Codex is the most production-ready option in that specific lane. For developers who want a closer, more interactive working relationship with their AI coding tool, Claude Code or Cursor will feel more natural.