gpt-5.4 and gpt-5.4-mini retire from the ChatGPT-subscription picker on 2026-08-31

[anandgupta42]

Summary

gpt-5.4 and gpt-5.4-mini are both in OAUTH_ALLOWED_MODELS (packages/opencode/src/plugin/codex.ts), and both retire from the ChatGPT-subscription model picker at 2026-08-31T19:00:00Z.

Evidence

openai/codex ships its model catalog in the repo. Fetched first-hand from https://raw.githubusercontent.com/openai/codex/main/codex-rs/models-manager/models.json:

gpt-5.4        visibility=hide  supported_in_api=True  retirement_at=2026-08-31T19:00:00Z
gpt-5.4-mini   visibility=hide  supported_in_api=True  retirement_at=2026-08-31T19:00:00Z

with explicit migration targets:

"gpt-5.4":      { "model": "gpt-5.6-terra", "migration_markdown": "GPT-5.4 is no longer available\n\nCodex now uses GPT-5.6 Terra in place of GPT-5.4. Switch to GPT-5.6 Terra to continue.\n", "retirement_at": "2026-08-31T19:00:00Z" }
"gpt-5.4-mini": { "model": "gpt-5.6-luna",  "migration_markdown": "GPT-5.4 Mini is no longer available\n\nCodex now uses GPT-5.6 Luna in place of GPT-5.4 Mini. Switch to GPT-5.6 Luna to continue.\n", "retirement_at": "2026-08-31T19:00:00Z" }

Scope — this is the subscription path only

This is not an API deprecation, and the two should not be conflated:

• both still carry supported_in_api: true
• neither appears on OpenAI's API deprecations page
• models.dev does not mark either deprecated (the 10 openai ids currently marked deprecated there are gpt-3.5-turbo, gpt-4, gpt-4-turbo, gpt-4.1-nano, gpt-4o-2024-05-13, gpt-image-1, o1, o1-pro, o3-mini, o4-mini — neither gpt-5.4 id is among them)

OAUTH_ALLOWED_MODELS is only consulted when auth.type === "oauth" (packages/opencode/src/plugin/codex.ts, the models loader returns {} for any other auth type), so API-key users are unaffected and keep both ids.

Why this will not fix itself

Because these are still live API models, models.dev will keep listing them — its stated policy is to hard-delete only once a model actually stops serving (see sst/models.dev#3707). So the catalog will not drop them for us.

After the deadline they would remain in the subscription picker and fail at request time with an opaque 400 — precisely the failure mode #1179 rebuilt this allowlist to prevent.

Fix

Remove both from OAUTH_ALLOWED_MODELS. Their documented replacements, gpt-5.6-terra and gpt-5.6-luna, are already in the set, so affected users land on a working model with no further change.

Timing matters and needs a human call. Removing before 2026-08-31T19:00:00Z breaks users who are successfully using gpt-5.4 today; leaving it after gives an opaque failure. See the linked PR for the reasoning on which way to err.

[anandgupta42]

Sequencing requirement, found in review of #1190.

The allowlist only ever deletes from the catalog — it cannot add a model the catalog lacks. Release binaries today embed a 2026-03-30 test fixture (that is #1186), and that fixture contains neither replacement model. So removing the two ids on its own has this effect:

catalog	allowed before	allowed after
shipped snapshot (cold cache today)	gpt-5.3-codex-spark, gpt-5.4, gpt-5.4-mini	gpt-5.3-codex-spark — one model
live models.dev	spark, 5.4, 5.4-mini, 5.5, sol, luna, terra	spark, 5.5, sol, luna, terra

So #1188 must ship before #1190. Once binaries embed a release-time catalog, gpt-5.6-terra and gpt-5.6-luna are present and the cold-cache regression does not occur.

Note that a runtime date gate would not have avoided this — the problem is a missing catalog entry, not a filter decision.