If Codex suddenly says GPT-5.4 is unavailable, that does not necessarily mean the model disappeared everywhere or that a usage limit was hit. OpenAI’s current guidance says that from August 31, 2026, GPT-5.4 and GPT-5.4 mini are no longer available in Codex when the session is authenticated with a ChatGPT account.

The important boundary is how Codex is authenticated.

OpenAI says the change does not affect the OpenAI API or Codex sessions using a personal API key. For ChatGPT-authenticated Codex, the recommended replacements are explicit:

  • GPT-5.4GPT-5.6 Terra
  • GPT-5.4 miniGPT-5.6 Luna

That makes this less like a full model shutdown and more like a product-surface migration. The places most likely to break are saved model selections, workspace defaults, managed configuration and automations that still request the old model names.

Status check: August 31, 2026. OpenAI’s Help Center was updated within the past day and still lists this retirement date and migration path.

What changes on August 31?

Where GPT-5.4 is usedWhat happensBest next step
Codex signed in with ChatGPTGPT-5.4 and GPT-5.4 mini are no longer availableMove to Terra or Luna
Codex using your own OpenAI API keyOpenAI says this August 31 change does not applyNo forced migration from this specific change
OpenAI API directlyThis August 31 Codex change does not remove the models from the APIContinue if needed, but keep monitoring model lifecycle notices
Saved Codex workspace/default configurationA pinned old model can become an unavailable selectionUpdate the stored model name
Scheduled or automated Codex workflowsJobs that explicitly request an old model can fail at model selectionReplace the model reference and test one run

The practical mistake is assuming that “GPT-5.4 retired” means the same thing in every OpenAI product. It does not.

Why authentication matters more than the model name

Codex can be used through a ChatGPT plan or with an API key. Those paths have different model-availability rules.

For this retirement, OpenAI draws the line clearly: ChatGPT sign-in loses GPT-5.4 and GPT-5.4 mini in Codex; API-key usage does not.

That creates a support problem that can look like a random bug inside a team. Two developers may open the same project, request the same model, and see different behavior simply because one Codex session uses ChatGPT authentication while the other uses an API key.

Before debugging anything deeper, answer these two questions:

  1. Which model is the workflow requesting?
  2. How is that Codex session authenticated?

If the answer is “GPT-5.4 through ChatGPT sign-in,” the August 31 change is the first thing to fix.

Terra vs Luna: which replacement makes sense?

OpenAI’s official mapping is simple, but the better choice still depends on what the old model was doing.

Choose GPT-5.6 Terra when quality and cost both matter

OpenAI describes GPT-5.6 Terra as the model that balances intelligence and cost. It is the direct replacement OpenAI names for GPT-5.4.

For most interactive coding work, that makes Terra the safest starting point when a workflow previously depended on the full GPT-5.4 model.

Typical cases include:

  • feature implementation;
  • debugging;
  • code review;
  • refactoring;
  • repository-wide changes;
  • multi-step engineering tasks where a very small model would be too brittle.

Choose GPT-5.6 Luna for cost-sensitive, high-volume work

OpenAI describes GPT-5.6 Luna as the cost-sensitive, high-volume option and explicitly maps GPT-5.4 mini to Luna.

That makes Luna a sensible fit for work such as:

  • repetitive transformations;
  • lightweight classification;
  • simple code edits;
  • low-risk sub-tasks;
  • workflows where the model is called many times and per-call quality requirements are lower.

Do not assume every Codex plan exposes every GPT-5.6 model

This is an easy detail to miss.

OpenAI’s current GPT-5.6 availability table says Codex on Free and Go currently includes Terra, while Plus, Pro, Business and Enterprise include Sol, Terra and Luna.

So the conceptual migration may be “mini → Luna,” but the actual model available in a specific Codex account can depend on the plan and workspace policy.

The safer rule is:

Use OpenAI’s recommended mapping first, then confirm that the replacement is actually available in the Codex plan and workspace where the workflow runs.

API price is not the same thing as ChatGPT Codex usage

The GPT-5.6 model pages currently list these API text-token prices:

ModelInput per 1M tokensCached inputOutput per 1M tokens
GPT-5.6 Terra$2.00$0.20$12.00
GPT-5.6 Luna$0.20$0.02$1.20

Those numbers are useful if a workflow pays through the OpenAI API.

They are not a reliable way to calculate how many Codex tasks a ChatGPT subscription gets. OpenAI says Codex usage limits vary by plan and depend on factors such as the model, task complexity, context, reasoning level, speed and tools used.

That distinction matters during migration. A team should not look at Luna’s API price and assume that its ChatGPT Codex allowance will scale by exactly the same ratio.

A five-minute migration audit

The fastest useful audit is not “open Codex and see whether the new model appears.” It is to find every place where the old model name might have been saved.

1. Check workspace defaults

OpenAI explicitly recommends updating workspace defaults before the cutoff.

For managed teams, also check whether an administrator has enforced a starting model or model requirement. A local preference cannot override an unavailable or centrally enforced model policy.

2. Search local configuration and project files

If model names have been written into scripts or configuration files, a quick text search can catch stale references:

rg -n 'gpt-5\.4(-mini)?' ~/.codex .

The exact storage location depends on how Codex and the project are configured, so treat this as a broad search rather than a complete inventory.

3. Check custom agents and reusable workflows

A model may not be visible in the main project configuration but can still be pinned inside an agent definition or reusable task.

Search the configuration for explicit model names rather than assuming the current interactive model controls every sub-task.

4. Check scheduled tasks and automations

OpenAI specifically names automations as something to update before August 31.

For each recurring workflow, record:

  • old model;
  • replacement model;
  • authentication method;
  • whether the task still completes successfully after the switch.

This turns a vague migration into a small compatibility matrix.

5. Test the replacement on representative work

Do not validate the migration with only a “hello world” prompt.

Pick two or three real tasks the old model handled frequently. For example:

  • one bug fix;
  • one multi-file edit;
  • one review or reasoning-heavy task.

Compare whether the replacement still follows repository instructions, uses tools correctly, finishes the task, and produces an acceptable diff.

The goal is not to prove Terra or Luna is universally “better.” The goal is to prove that the new model works for this workflow.

Should anyone keep GPT-5.4 through an API key?

Possibly, but mainly as a compatibility bridge.

OpenAI says this August 31 change does not affect API-key Codex sessions or the OpenAI API, so a workflow that genuinely depends on GPT-5.4 can continue through that route for now.

But switching authentication just to avoid one product-surface retirement creates new responsibilities:

  • API billing;
  • key management;
  • usage controls;
  • a separate model-lifecycle dependency.

That can be reasonable for production automation or a carefully controlled migration. It is usually less attractive as a permanent workaround for an interactive workflow that can simply move to Terra.

What is confirmed, and what is not

Confirmed by OpenAI as of August 31

  • GPT-5.4 and GPT-5.4 mini stop being available in ChatGPT-authenticated Codex on August 31, 2026.
  • GPT-5.4 should be replaced with GPT-5.6 Terra.
  • GPT-5.4 mini should be replaced with GPT-5.6 Luna.
  • Workspace defaults, saved model settings, managed configurations and automations should be updated.
  • The August 31 change does not affect the OpenAI API or Codex sessions using a personal API key.
  • OpenAI currently describes Terra as the balance of intelligence and cost, and Luna as the cost-sensitive, high-volume option.

Not established by this retirement notice

  • that a special Codex CLI update is required specifically for this model removal;
  • that API access to GPT-5.4 will remain available indefinitely;
  • that every ChatGPT plan or managed workspace exposes the same GPT-5.6 choices;
  • that changing only the top-level model guarantees every custom agent or automation has migrated.

Those are the areas where checking the actual runtime configuration is more useful than relying on the model picker alone.

What to watch next

Three things are worth monitoring after the cutoff.

1. Hidden stale model references

The obvious settings will be changed first. The failures that appear later are more likely to come from an old automation, custom agent or workspace policy that nobody remembered was pinned to GPT-5.4.

2. Plan-specific model availability

OpenAI’s current availability table differs by plan. If a documented replacement does not appear in Codex, check the plan and workspace policy before assuming the rollout is broken.

3. The separate API lifecycle

The August 31 notice is scoped to ChatGPT-authenticated Codex. API users should follow OpenAI’s model documentation and deprecation notices separately rather than treating this date as the API shutdown date.

Conclusion

The useful way to understand the August 31 GPT-5.4 change is not “OpenAI deleted GPT-5.4.”

It is:

ChatGPT-authenticated Codex no longer offers GPT-5.4 or GPT-5.4 mini, while API-key Codex and the OpenAI API are outside this specific retirement.

For most users, the migration is straightforward: move GPT-5.4 workflows to GPT-5.6 Terra, move mini-style workloads to GPT-5.6 Luna where the plan supports it, then check saved configuration and automations for hidden old model references.

The concrete next step is a small inventory: model name + authentication method + where that model is saved. That catches the majority of avoidable failures before they become mysterious Codex errors.

Sources

Checked August 31, 2026:

Written and reviewed by /lico

Just writing down my thoughts, interests, and the things I learn along the way.