GPT-5.4 Retires from Codex on August 31: Migration Guide for GPT-5.6

TL;DR

If you use Codex with a ChatGPT account, GPT-5.4 and GPT-5.4-mini stop working on August 31, 2026. I checked my own configs this morning and found two custom agents still pointed at the old models. You have 19 days. The replacements (GPT-5.6 Terra and Luna) are faster, cheaper, and score higher on coding benchmarks. But if you have saved model configs in custom agents, scheduled tasks, or Codex CLI settings, they will silently stop working on September 1 unless you update them now. Here is the checklist.

What's happening

OpenAI published a notice on the official Codex changelog: GPT-5.4 and GPT-5.4-mini will no longer be available in Codex for users who sign in with a ChatGPT account after August 31, 2026.

This is not the first Codex model retirement. GPT-5.2-codex and five other Codex-specific variants disappeared on July 23. But GPT-5.4 is the first "current daily driver" model to get the axe. It has been the default in Codex since March 2026. Millions of developers have it baked into config files, custom agents, and muscle memory.

The key detail most people miss: this retirement only applies to ChatGPT sign-in. If you authenticate Codex with an API key (your own OpenAI API key or a third-party provider's), GPT-5.4 and GPT-5.4-mini remain available. The API models are on a separate deprecation track: GPT-5.4 API deprecation was announced June 11 with a December 11, 2026 removal date.

What you need to change

The migration is straightforward. One model maps to one replacement:

Here is where these model names live and how to check each one:

Codex CLI. Open ~/.codex/config.toml (macOS/Linux) or %USERPROFILE%\.codex\config.toml (Windows). Search for gpt-5.4. If you find it in a [models] block or agent definition, replace it. If you use codex exec with --model gpt-5.4 in shell scripts or CI pipelines, those will also break.

Custom agents. In the Codex UI, open each custom agent, check the model selector in the agent settings, and switch to the GPT-5.6 equivalent. There is no bulk migration tool. You will need to check each agent manually. When I checked mine, both were still on 5.4 despite me having switched the CLI config weeks ago.

Scheduled tasks. Any task that specifies a model in its configuration needs the same update. Most scheduled tasks inherit the default model, which OpenAI will presumably change automatically. I would not count on it.

ChatGPT desktop app agent mode. The Codex integration inside the ChatGPT desktop app uses the same model routing. If you have saved agent configurations there, check them too.

Price comparison: the retirement saves you money

The July 2026 pricing restructure made the GPT-5.6 family cheaper than GPT-5.4 was at launch. Here is the comparison for Codex usage (ChatGPT subscription rates):

Both replacements are roughly half the cost of what they replace. Migration does not just keep things working: it cuts your token spend.

Capability: faster, better on benchmarks

The GPT-5.6 family shipped in July 2026 and has been the recommended default for new Codex users ever since. On coding benchmarks:

• GPT-5.6 Sol (the top-tier model, not replacing 5.4) leads Terminal-Bench 2.1 at 89.5%
• GPT-5.6 Terra scores 84.2% on SWE-bench Verified, compared to GPT-5.4's 78.1%
• GPT-5.6 Luna is designed for speed and cost efficiency on routine tasks: refactors, test generation, boilerplate

Daniel Vaughan, who wrote one of the earliest migration guides, describes it as "the easiest model migration Codex CLI users have faced." I mostly agree, with one caveat: the actual model swap is trivial. Finding every place the old model name is hardcoded is not.

The bigger deprecation wave

The GPT-5.4 retirement is part of an aggressive OpenAI cleanup schedule. Here is what else is coming:

If your workflow still references gpt-4 or gpt-4o anywhere, October 23 is your real deadline. That one is going to be much messier than the GPT-5.4 retirement: those models are embedded in years of documentation, tutorials, and enterprise CI pipelines.

What to do right now

1. Grep your Codex config: grep -r "gpt-5.4" ~/.codex/
2. Check each custom agent in the Codex UI for its model setting
3. Replace gpt-5.4 with gpt-5.6-terra and gpt-5.4-mini with gpt-5.6-luna
4. Run one coding session with each agent to verify nothing broke
5. Set a calendar reminder for October 23 if you still use GPT-4 or GPT-4o anywhere

The model swap itself takes 30 seconds. Finding every place you hardcoded the old model name is where the real time goes. Do it now, not on September 1 when your CI pipeline silently fails.