GPT-5.4 and GPT-5.4 mini are now retired from Codex for ChatGPT-authenticated users
From 31 August 2026, GPT-5.4 and GPT-5.4 mini no longer work in Codex under ChatGPT sign-in. Here is what to do.
The deadline is live
From 31 August 2026, gpt-5.4 and gpt-5.4-mini are no longer available in OpenAI Codex for anyone signed in with a ChatGPT account. OpenAI announced the cutoff on 31 July 2026 and gave one month’s notice. That month is now up.
If your Codex workflow uses ChatGPT authentication and has either model ID hard-coded anywhere, it is broken right now. There is no automatic remap, no silent fallback to a working model, and no grace period beyond the announced date.
Who is affected and who is not
This is an access-path change, not a model shutdown. The distinction matters.
Affected: Any Codex session authenticated via a ChatGPT account (Plus, Team, Enterprise, or free tier) that explicitly selects gpt-5.4 or gpt-5.4-mini. This includes the Codex CLI, the IDE extension, the Codex desktop and web app, custom agents built on those surfaces, and any scheduled task that fires against a model picker tied to ChatGPT sign-in.
Not affected: Codex sessions authenticated with an OpenAI API key, and direct calls to the OpenAI API itself. If you are routing requests through the Responses API with an API key, gpt-5.4 and gpt-5.4-mini remain available on that path for now.
The practical upshot: teams using Codex as a product, signed in the same way they sign into ChatGPT, need to act. Teams using Codex as an API surface with their own key do not, at least not because of this announcement.
What to switch to
OpenAI’s recommended replacements are straightforward:
- Replace
gpt-5.4withgpt-5.6-terra(GPT-5.6 Terra) - Replace
gpt-5.4-miniwithgpt-5.6-luna(GPT-5.6 Luna)
GPT-5.6 launched on 9 July 2026 after a limited preview and comes in three tiers. Sol is the flagship. Terra sits in the middle, with performance competitive with GPT-5.5 and a price roughly half that of its predecessor. Luna is the fastest and most affordable tier.
On 30 July 2026, OpenAI cut the price of GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20%, so the successor models are meaningfully cheaper than what you were using before:
| Model | Input per 1M tokens | Output per 1M tokens |
|---|---|---|
| GPT-5.6 Sol | $5.00 | $30.00 |
| GPT-5.6 Terra | $2.50 | $15.00 |
| GPT-5.6 Luna | $1.00 | $6.00 |
Terra is the natural like-for-like replacement for gpt-5.4: capable, cost-effective, and intended for the same kind of coding and reasoning workloads. Luna replaces gpt-5.4-mini for tasks where speed and low cost matter more than raw capability.
Where to look for broken configurations
The failure modes are specific enough to be worth listing individually. Check each of these before assuming your setup is clean:
Workspace defaults. If your team or personal Codex workspace was set to use gpt-5.4 or gpt-5.4-mini as the default, those sessions will now fail to start or fall into an error state. Update the workspace model setting to gpt-5.6-terra or gpt-5.6-luna.
CLI and IDE configs. Any .codex/config file, VS Code settings, or JetBrains plugin configuration that specifies a model ID explicitly needs updating. A global search for gpt-5.4 across your dotfiles and project configs is the quickest way to find these.
Managed configurations. If your organisation manages Codex settings centrally, for example through a shared config file or a Codex Teams admin panel, the model ID in that configuration needs updating. Individual users will not be able to override a managed config that points to a retired model.
Custom agents. Any agent built on Codex that hard-codes gpt-5.4 or gpt-5.4-mini in its system prompt scaffold, tool definition, or routing logic will fail when invoked under ChatGPT authentication.
Scheduled tasks. Recurring Codex tasks, such as nightly code review runs or automated PR summaries, will start throwing errors if they reference either retired model ID. Update the model parameter in the task definition before the next scheduled run.
A note on credits and quotas
Unused ChatGPT plan credits or workspace quota do not restore a model that has been removed from the ChatGPT-authenticated Codex catalogue. Quota exhaustion means a meter hit zero on a model that is still listed. This announcement changes the listing itself. No amount of remaining credit brings back a retired model ID on this access path.
The broader picture
This retirement sits alongside a busy period of changes to OpenAI’s developer surfaces. The Assistants API was shut down on 26 August 2026, with /v1/assistants, /v1/threads, and /v1/threads/runs now returning hard errors. Developers still on that API need to migrate to the Responses API and Conversations API independently, as no automated migration tool exists.
Looking further ahead, o3 and o3-pro API snapshots are scheduled to deprecate on 11 December 2026. The recommended migration path for those is gpt-5.6-sol, with reasoning.mode: pro for workloads that currently use o3-pro.
None of that is today’s problem. Today’s problem, if you use ChatGPT-authenticated Codex with either retired model ID, is the one described above. The fix is a find-and-replace on a model string, and the replacement models are cheaper and more capable than what they are replacing. It is a straightforward migration once you know where to look.