OpenAI Codex release notes: replace GPT-5.4 by August 31 and review across repos

OpenAI has set an end date for two models in Codex. On August 31, 2026, GPT-5.4 and GPT-5.4 mini leave Codex for users who sign in with ChatGPT. The day before that notice, the ChatGPT desktop app gained cross-repository review for local projects with multiple folders. Together, the changes give teams a concrete August task: find old model selections, replace them, and review the migration as one coherent change set.

OpenAI Codex release notes: the model migration and cross-repository review

OpenAI Codex is a coding agent that can read, change, and test code in the terminal, IDE, cloud, and ChatGPT desktop app. A multi-folder project is a local project where several related folders or repositories share one workspace.

OpenAI published two relevant entries in the Codex changelog:

• July 31: GPT-5.4 and GPT-5.4 mini receive an end date in Codex with ChatGPT sign-in.
• July 30, desktop app 26.727: the Review view can show changes and diffs across every repository in a multi-folder project.

No newer Codex CLI build has appeared after 0.147.0-alpha.4. npm still points to 0.146.0 as latest and 0.147.0-alpha.4 as alpha. Today's change comes from the product changelog rather than a new CLI tag.

Source: OpenAI Codex changelog entries for July 30 and 31.

Source: npm registry metadata for @openai/codex.

GPT-5.4 gets two explicit Codex replacements

For Codex sessions that use ChatGPT sign-in, OpenAI specifies these replacements:

• Replace gpt-5.4 with gpt-5.6-terra.
• Replace gpt-5.4-mini with gpt-5.6-luna.

OpenAI names workspace defaults, saved model settings, managed configurations, custom agents, and scheduled tasks as places to check before August 31. The retirement does not affect the OpenAI API or Codex sessions authenticated with an API key.

That distinction matters when one organization uses several Codex surfaces. A model name in a ChatGPT-connected scheduled task may need to change while an API integration can keep using the model. The inventory should therefore record both the model reference and the run's authentication method.

Source: OpenAI Codex models and deprecated models.

Source: OpenAI on workspace model access and the GPT-5.4 migration.

Desktop app 26.727 brings several repository diffs into one review

In a local multi-folder project, Codex can read and change files in every attached folder. The primary folder still controls new chats, Git operations, and automatic discovery of AGENTS.md, skills, and config.toml. Secondary folders remain available for file search, reading, and editing.

Version 26.727 lets the Review view show the repositories and changed lines in each one. The user selects Review to inspect those diffs together instead of switching between separate review views. That is useful when a model migration touches an app, an automation repository, and its documentation at the same time.

Source: OpenAI Codex changelog for desktop app 26.727.

Source: OpenAI on local projects with multiple folders.

What Nordic automation teams should test now

Start with a read-only inventory. Search for the two old model names in repositories and versioned configuration. Add workspace defaults, custom agents, and scheduled tasks that live outside the repository to the same checklist.

Each match needs four details: where the reference lives, who owns it, whether the run uses ChatGPT sign-in or an API key, and which test will prove the replacement works. Make replacements only after confirming the authentication boundary. Then review the diffs across every affected repository and run the existing tests.

Keep API keys in environment variables or a secret manager. Record the model selection and authentication method in the run log so the team can distinguish an intentional API run from a ChatGPT-connected task after August 31.

Human step

1. Open the project's menu in the ChatGPT desktop app and select Edit project.
2. Select Add folder to attach related folders. Use Make primary on the folder that should control new chats, Git operations, and automatic instruction discovery.
3. Run the prompt below before changing any model reference.
4. After the changes have been reviewed by a person, select Review to inspect diffs across the attached repositories.

Source: OpenAI on the verified project menu, Add folder, Make primary, and Review view.

Short example: use the new Codex feature

Inventory every attached repository for exact references to gpt-5.4 and gpt-5.4-mini. Report file and line, owner or scope, authentication method when evidence supports it, and the recommended replacement with gpt-5.6-terra or gpt-5.6-luna. Also list settings outside the repository that a person must check. Change nothing.

Good output should:

• Show real files and lines or state clearly that no match was found.
• Separate ChatGPT sign-in from API-key authentication without guessing.
• Map each retired model to the correct replacement.
• Leave files, settings, and scheduled tasks unchanged.

What should you watch next?

The operational checkpoint is August 31. Watch the next stable Codex CLI release as well in case it changes model selection or multi-folder project behavior. Until then, teams can treat this as a normal reviewed configuration migration with an owner and a test result.

If model selections are scattered across several workflows, Tool Forge can collect them in a versioned model policy with ownership and run logging.