GPT-5.5 Leaves Codex on October 14

The mistake

Leaving gpt-5.5 in a Codex config, profile, script or workspace default past October 14, 2026, on the assumption that a model that worked yesterday will keep working — or that Codex will quietly swap in something similar.

Why this happens

GPT-5.5 was the flagship until recently, so it's baked into a lot of setups: a model line written months ago, a profile for long refactors, a scheduled task, a codex exec -m gpt-5.5 in CI. OpenAI is retiring it from ChatGPT, ChatGPT Work and Codex on every plan on that date. The API is not affected, which adds to the confusion — the same model id keeps working for anyone calling it with an API key, so "it works for me" can be true and irrelevant at the same time.

Why it matters

OpenAI doesn't describe a fallback for a retired model, so any session, agent or job that still selects GPT-5.5 through ChatGPT sign-in should be treated as broken from the retirement date. Scheduled and unattended work is the likeliest to fail unnoticed, because nobody is watching when it starts. And the fix isn't a single file: OpenAI lists workspace defaults, saved model settings, managed configurations, custom agents, scheduled tasks and scripts as places to update.

The fix

Replace gpt-5.5 with the successor OpenAI names for your plan: GPT-6 Sol (gpt-6-sol) on Plus, Pro, Business, Enterprise and Edu, or GPT-6 Luna (gpt-6-luna) in the desktop app on Free and Go. In Enterprise workspaces, an administrator has to enable the replacement first — changing a default doesn't grant access. If you'd rather follow the CLI's default, GPT-6.1 Sol, removing the model line works too. Then search configs, repositories and CI definitions for any remaining gpt-5.5, and confirm the active model with /status after the change.

If you call Codex with your own API key, nothing here forces a change; GPT-5.5 remains available on the API.

Checking after the change

Run /status in a fresh session to confirm the model, then trigger each scheduled task or CI job once by hand rather than waiting for its next run. Unattended work is where a stale model id hides longest, because nothing fails until nobody is watching.