Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- From August 31, 2026, Codex signed in with a ChatGPT account no longer offers GPT-5.4 or GPT-5.4 mini.
- The official mapping is GPT-5.4 to GPT-5.6 Terra and GPT-5.4 mini to GPT-5.6 Luna.
- The OpenAI API and Codex using your own API key are unaffected by this retirement. Do not indiscriminately replace API workloads.
- Check workspace defaults, saved model settings, managed configurations and automations.
- The notice does not explain whether old settings are rewritten automatically, how running sessions are handled or exactly when each client switches. Validate migration task by task.
A practical failure: a scheduled task suddenly cannot find its model
Interactive Codex still opens, but a team's overnight automation fails at startup with an unavailable or missing old-model error, while another member sees no problem. The cost extends beyond manually reselecting a model: unattended tasks stop, managed settings override personal choices, and replacement models may change task behavior without review.
An unhelpful first response is upgrading the client or globally replacing every GPT-5.4 string. First determine whether the failing task uses ChatGPT authentication or a personal API key, then record the source of the effective model setting. Authentication determines whether this change applies.
The retirement covers Codex signed in with a ChatGPT account
OpenAI's official help page states that from August 31, 2026, GPT-5.4 and GPT-5.4 mini are no longer available in Codex when signed in with a ChatGPT account. Its migration mapping is GPT-5.4 to GPT-5.6 Terra and GPT-5.4 mini to GPT-5.6 Luna.
This does not affect the OpenAI API or Codex using your own API key. Do not describe a model retirement for the Codex sign-in path as a universal API shutdown. Changes to API workloads require their own lifecycle and compatibility evidence.
Use authentication and configuration inventories to find old references
Group workloads by authentication first. Codex using ChatGPT sign-in is in scope; initially retain configuration for OpenAI API or personal-API-key workloads. Then check workspace defaults, organization/admin managed configurations, personal saved model settings, and explicit model names in agents, scripts and automations in priority order.
If the UI shows a new model but a task still uses the old one, inspect higher-priority workspace or managed settings instead of repeatedly reinstalling the client. When one repository is used with multiple authentication methods, record each method, model source, owner and validation result separately. Avoid replacing two paths with different lifecycles in one operation.
A six-step migration and validation checklist
First, export old model references and task owners. Second, mark each authentication method. Third, select Terra or Luna according to the official mapping only for ChatGPT-authenticated Codex. Fourth, update workspace, managed configuration, saved setting and automation references in sequence. Fifth, validate tool calls, permissions, output formats and completion times with small reproducible tasks. Sixth, restore unattended runs in batches and keep snapshots of prior configuration.
Acceptance is not merely seeing a new name in a model picker. Each critical task must start with the expected authentication, load the intended configuration, complete representative steps and expose whether a failure concerns the model, permissions or networking. The API key, base URL and model configuration guide helps distinguish configuration fields, but API field descriptions do not define the scope of this retirement.
Failure, rollback and stop conditions
Pause a task if managed configuration keeps overriding local values, the replacement behaves differently in critical tool calls or approvals, or an unattended task cannot be reproduced reliably. Do not expand bulk replacement. Restoring a saved configuration snapshot can help isolate a fault, but a retired model cannot be a long-term rollback target on the ChatGPT sign-in path.
The official notice does not specify automatic rewriting of old settings, handling of running sessions or exact switch times by client or region. Keep those points UNKNOWN and revisit the official help page. A temporary state in one client does not prove that every account has or has not switched.
Common mistakes and the network boundary
Common mistakes include forcing API-key users to migrate, changing personal settings but missing admin configuration, treating Terra and Luna as untested drop-in equivalents, and immediately restarting every scheduled task. This notice provides migration mappings, not guarantees of capabilities, pricing, quotas or synchronized switching across clients.
An unavailable model is primarily a product and configuration issue, not an egress-IP issue. Use the proxy connection checklist only with separate evidence such as DNS, TLS, 407, timeout or server-connectivity failures. Changing networks will not restore a retired model to Codex signed in with ChatGPT.
When to recheck and where to record later facts
Check all unattended tasks on migration day, review failure logs after the next scheduled run, and revalidate after an administrator changes workspace or managed configuration. If OpenAI later clarifies client behavior, automatic rewriting, regional or plan rules, or running sessions, update this page rather than creating overlapping pages about missing models, configuration errors and automation failures.
Record authentication, effective model source, client version, task start time, exact error and recovery action during review. Without new official facts, do not describe one account's picker state as a global change in availability.
Sources
Frequently Asked Questions
When was GPT-5.4 retired from Codex?
OpenAI states that from August 31, 2026, GPT-5.4 and GPT-5.4 mini are unavailable in Codex signed in with a ChatGPT account.
What replaces GPT-5.4 and GPT-5.4 mini?
The official mappings are GPT-5.4 to GPT-5.6 Terra and GPT-5.4 mini to GPT-5.6 Luna.
Was GPT-5.4 also retired from the OpenAI API at the same time?
This notice explicitly does not affect the OpenAI API or Codex using your own API key. Do not conflate the two lifecycles.
Why does automation still report the old model after I changed it?
Check explicit references in workspace defaults, managed configurations, saved settings and automations. Higher-priority configuration can override a personal selection.
Will old configuration automatically become Terra or Luna?
The notice does not specify automatic rewriting. Treat it as unknown and inspect and validate each setting rather than relying on automatic migration.
Can changing IPs fix an unavailable model?
There is no evidence that changing an IP restores a retired model. Investigate networking only when separate DNS, TLS, 407 or timeout evidence exists.