OpenAI put a retirement date on GPT-5.4 in the Codex changelog on July 31, and it lands four weeks from today. If gpt-5.4 or gpt-5.4-mini is written anywhere in your Codex setup, that name stops resolving on August 31.
What OpenAI actually said
The changelog entry reads: “On August 31, 2026, GPT-5.4 and GPT-5.4 mini will no longer be available in Codex for users signed in with ChatGPT.”
How you authenticate is the entire scope here. OpenAI says both models remain available on the OpenAI API and in Codex sessions authenticated with an API key. The cutoff only reaches you if Codex is signed in through a ChatGPT login.
What replaces what
OpenAI names a direct swap for each model:
gpt-5.4→gpt-5.6-terragpt-5.4-mini→gpt-5.6-luna
The Codex models reference calls Terra the balanced GPT-5.6 model for everyday work, with performance competitive with GPT-5.5 at a lower cost. Luna is listed as the fast, affordable option with the lowest cost in the family. Sol sits above both as the flagship, described as the strongest for complex coding, computer use, research, and cybersecurity.
If Sol or Terra is already your daily driver, the deadline doesn’t touch your default. It hits the spots where you pinned gpt-5.4 months ago and stopped thinking about it. GPT-5.4 Thinking and Pro shipped March 5, 2026, and gpt-5.4-mini followed on March 17, so anything you configured in the spring is worth opening.
Mini is easy to overlook because you may never have chosen it on purpose. OpenAI’s docs describe it as a fast, efficient mini model for responsive coding tasks and subagents.
The five places to check
OpenAI’s instruction is specific: “Before the cutoff, update workspace defaults, saved model settings, managed configurations, custom agents, and scheduled tasks that use either model.”
Broken out, that’s five separate surfaces:
- Your workspace default model setting
- Saved model settings in your own Codex config
- Managed configurations, if you’re on a team plan where an admin pushes settings down
- Custom agents with a model name hardcoded
- Scheduled tasks, including cron jobs and anything that runs Codex unattended
Where this bites
A model you pick from a dropdown is easy to catch going stale, because you see it every time you open a session. A model name sitting in a scheduled task’s config file, written once in the spring and never opened since, gives you nothing to notice. Same story with a custom agent somebody on the team built in March that nobody has needed to touch.
OpenAI’s changelog doesn’t say what happens to a request that still names gpt-5.4 after August 31, so I can’t tell you whether it errors out or gets swapped to something else. Either way, a nightly job on September 1 is a bad place to find out.
Craig’s take
Four weeks is fair notice, and it’s also the kind of deadline that only reaches the person who reads changelogs. If you run Codex on a schedule, in CI hooks, or inside custom agents wired into other tools, grep your configs for gpt-5.4 and gpt-5.4-mini this week rather than on August 30. The swap itself is a one-line change; finding every place the name got written is the part that takes an afternoon.
For what each GPT-5.6 model is built for, see our models reference.