Guide

GPT-5.4 Leaves ChatGPT-Signed Codex on August 31. The API Is Not a Quota Refill

By Istar editors
At a glance

ChatGPT-login Codex loses GPT-5.4 and GPT-5.4 mini on August 31, while API access remains. Decide whether to migrate, switch, or remove the pin.

If your Codex workflows still name gpt-5.4 or gpt-5.4-mini under a ChatGPT login, August 31 is a product cutoff. It is not a usage reset.

On 31 July 2026, the ChatGPT & Codex changelog said the part that matters this week: on 31 August 2026, GPT-5.4 and GPT-5.4 mini will no longer be available in Codex for users signed in with ChatGPT. The same note says both models remain available on the OpenAI API and in Codex sessions authenticated with an API key. OpenAI’s recommended replacements are explicit: replace gpt-5.4 with gpt-5.6-terra, and replace gpt-5.4-mini with gpt-5.6-luna. Before the cutoff, update workspace defaults, saved model settings, managed configurations, custom agents, and scheduled tasks that use either model.

That is an authentication-path split. ChatGPT-signed Codex loses two model IDs. API-key Codex and the API keep them. Unused ChatGPT plan or workspace credits do not restore a model removed from the ChatGPT-signed Codex catalog. Quota exhaustion is a meter hitting zero on a still-listed model; this announcement changes the listing. After August 31, if ChatGPT-signed Codex reports that 5.4 is unavailable, check the authentication path and model catalog before treating it as quota exhaustion.

Source: ChatGPT & Codex changelog, 31 July 2026.

What actually breaks on Monday

The changelog does not say ChatGPT chat is losing 5.4. It does not say the API is sunsetting 5.4. It does not say Codex as a product is going away. It says Codex for users signed in with ChatGPT will no longer offer GPT-5.4 and GPT-5.4 mini. Confirm the current catalog before acting.

If you are pinned, the failure mode is local and specific:

  • Workspace default still set to gpt-5.4 or gpt-5.4-mini.
  • Saved model settings in CLI, IDE, or the Codex app.
  • Managed configuration that locks a model ID for a team.
  • Custom agents that hard-code the old IDs.
  • Scheduled tasks that keep firing against a picker that no longer includes them.

OpenAI tells users to update those surfaces before the cutoff. It does not promise an automatic remap, a silent fallback, or a grace week after August 31. Do not budget for a behavior the changelog does not describe.

If you are not pinned — if you already accept the current recommended model — this is a notice, not a fire drill. The people who need a decision are the ones who tuned prompts to 5.4’s tone, regress against 5.4 goldens, or have a policy that names a model version.

You have three real options. Only two keep a named model under your control.

Keep 5.4. Move the pinned work onto the OpenAI API or Codex sessions authenticated with an API key. Those are the first-party paths documented in the changelog for retaining these model IDs. You keep the model string but leave the ChatGPT-login bundle. Billing moves to the OpenAI Platform account, API-organization retention and data-sharing settings apply, and key management plus usage monitoring become part of the migration. OpenAI’s Codex authentication guide documents the billing and governance split.

Switch. Follow OpenAI’s map: 5.4 → GPT-5.6 Terra, and 5.4 mini → GPT-5.6 Luna. You can stay on ChatGPT-signed Codex if that is how you work today. You accept a new model family, and you owe yourself an evaluation — not a hope that “recommended” means “same.”

Do nothing. This is rational only if nothing in production names these IDs. If something does, doing nothing schedules a break.

Do not invent a fourth option in which ChatGPT-login Codex keeps 5.4 because the workspace still has credits. Workspace usage limits control consumption; they do not put a removed model back into the catalog.

How to decide in the three-day window

Start with why you pinned, not with price.

Stay and migrate if any of these are true:

  • A compliance, customer, or internal change-control document names gpt-5.4 or gpt-5.4-mini. A recommended successor is still a model change.
  • You have a golden set, prompt pack, or style contract that 5.4 already passes and you cannot re-baseline before Monday.
  • Downstream systems parse 5.4’s shape — patch style, test layout, refusal tone — and a miss is a production incident, not merely a worse coding session.
  • You already run API-key Codex or API calls for the same work, so the new path mostly points pinned jobs at an auth method you already operate.

Switch if these are true:

  • The pin was habit, a leftover default, or “this one was good in April,” not a locked baseline.
  • You can run Terra or Luna against the same tasks before August 31 and accept or reject on evidence.
  • You want to keep ChatGPT-signed Codex and will not stand up API billing, key hygiene, and spend caps this weekend.

Price is a modifier, not the first filter. On the OpenAI API pricing tables retrieved 28 August 2026, both 5.4 IDs were still listed and priced. They were not in the HTML “flagship / our latest models” grid, which showed gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna; they were in the full standard, batch, flex, and fast tables. Listed for API billing is not the same as featured. Confirm the live price before moving work.

Standard short-context prices per 1 million tokens on that date were:

Model Input Cached input Cache writes Output
gpt-5.4 (<272K) $2.50 $0.25 — $15.00
gpt-5.4-mini $0.75 $0.075 — $4.50
gpt-5.6-terra $2.00 $0.20 $2.50 $12.00
gpt-5.6-luna $0.20 $0.02 $0.25 $1.20

Long-context rows list gpt-5.4 at $5.00 input, $0.50 cached input, and $22.50 output; Terra also has long-context rates. Mini has no long-context prices in that table. Batch and flex are lower, while fast is higher. Regional processing adds a 10% uplift for eligible models released on or after 5 March 2026; check whether the chosen endpoint qualifies before copying that line into a forecast.

Source: OpenAI API pricing, including the complete pricing table.

A worked illustration, not a forecast: suppose a pinned 5.4 job uses 20 million input and 5 million output tokens a month at standard short-context rates, with no cache. 5.4 is about $125; Terra is about $100 before cache writes. The same volume on mini is about $37.50, versus about $10 on Luna. If the work is currently inside ChatGPT-signed Codex, that API figure is new usage-billed cash rather than a replacement line on the subscription invoice. If the work already uses the API, the comparison is model versus model. Recalculate from the live table: token mix, cache hit rate, long context, batch, and fast mode can move the total more than this illustration does.

Migration cost is more than the token row

Keeping 5.4 is not free even when the model ID stays.

Inventory. Count every pin OpenAI named: defaults, saved settings, managed configuration, custom agents, and scheduled tasks. One forgotten scheduled job is an incident.

Auth and money. ChatGPT login and API keys are different doors. API-key Codex is usage-billed on the Platform account. Establish a key policy, spend controls, and usage monitoring, with a named owner for the invoice. That operating cost can exceed a month of 5.4 tokens if you do not already manage API auth.

Data and admin. Moving off ChatGPT-signed sessions changes the billing account and applies API-organization retention and data-sharing settings; some ChatGPT-workspace or cloud-dependent features may also differ. If you pinned 5.4 for compliance, the authentication change is part of the change record. Do not assume “same model ID” means “same processing path.”

Switching cost. If you take Terra or Luna, budget hours rather than vibes: replay a fixed task pack, compare diffs and tests, watch for style drift, then retune prompts only where the new model fails. A weekend bakeoff is enough to reject a swap. It is rarely enough to rebuild a locked baseline. If you cannot finish that loop before August 31 and must keep the old ID, the API or API-key path is the documented first-party route.

Doing nothing. It costs no work now. After the cutoff, ChatGPT-signed Codex no longer lists those IDs. The changelog does not document a courtesy redirect. Price that as a possible hard failure on every still-pinned surface.

What this is not

It is not “5.4 is deleted.” The API pricing tables still carried gpt-5.4 and gpt-5.4-mini when checked on 28 August.

It is not “your Codex allowance ran out.” Allowance is a meter. This is a catalog change for one login type.

It is not a promise that API 5.4 is forever. Remaining available on August 31 is not a deprecation date. If you migrate to pin the ID, watch the API model list the same way you had to watch Codex.

Decision

  • Move and keep 5.4: the pin is load-bearing — compliance, goldens, or an output contract — and you can stand up API or API-key Codex before August 31.
  • Switch to Terra or Luna: the pin is not load-bearing, you can evaluate the recommended replacement against real tasks this weekend, and you want to stay on ChatGPT-signed Codex.
  • Do nothing: no workspace default, saved setting, managed configuration, custom agent, or scheduled task still names gpt-5.4 or gpt-5.4-mini.