Why OpenAI scheduled it for shutdown
Codex-style variants are specialized snapshots that OpenAI folds back into the main line as the flagship's own coding ability catches up. When that happens the specialized id is retired on a date like the one above, and calls that name gpt-5-codex stop resolving.
Because this model is used inside developer tooling and agents — often on a code path that runs unattended — a shutdown tends to surface as a broken build step or a stalled agent rather than an obvious user-facing error.
What to move to
The replacement above is the current flagship, whose general coding performance is intended to cover what the dedicated Codex snapshot did. For most code generation and review tasks it is a direct substitute.
If you drive an agent or IDE integration that has gpt-5-codex written into its configuration, that config is the thing to update — the tool won't pick a new model for you.
Migration notes
Coding workloads are sensitive to output shape: diff formats, tool-call conventions, and how strictly the model stays inside a requested patch. Re-run your existing eval or a representative batch of tasks against the replacement before cutting over, rather than assuming the prompt transfers unchanged.
Search your repos and tool configs for the literal gpt-5-codex, including CI files, agent manifests, and any .env that holds a model name — these are the spots a plain grep of application code can miss.