OpenAI released Codex — a cloud assistant for programmers.
It can work with repositories, run tests, and create Pull Requests. Codex is available in ChatGPT (Pro, Team, Enterprise; Plus and Edu coming soon) and as a CLI utility with a lightweight model.

Key features:

  • Clones a repo, edits code, opens a PR.
  • Works in isolated containers.
  • Runs tests and linters on its own and fixes code until it passes.

Technically:

  • Codex-1 — a fine-tuned version of o3, up to 192k tokens.
  • Codex-mini — a model based on o4-mini for CLI and simple tasks.

Metrics:

  • ~75% accuracy on SWE-Bench.
  • Writes hundreds of lines of code, passes tests, makes PRs — you can stay out of it.

Interface:

  • In ChatGPT — a panel with Code and Ask tabs.
  • CLI — local work with codex-mini, sign-in via ChatGPT, automatic setup.

Pricing:

  • Temporary free access for Pro/Team/Enterprise.
  • CLI: $1.50 per 1M input tokens, $6 per 1M output, 75% discount with caching.

Cons:

  • The Codex name conflicts with another tool.
  • UI in ChatGPT can be distracting and is limited in features.
  • Few integrations with IDEs and trackers.

Security:

  • Containers without internet.
  • Agent logs and steps are visible.
  • Measures against malicious requests.

Plans:

  • Expanding access and integrations.
  • Improving the interface and agent interaction.
  • Merging the copilot and task delegation modes.

Codex is a step toward automating programming:
an agent that works with code and tests, but still needs improvements for mass adoption.