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
- OpenAI's current deprecations page confirms that API access to the listed legacy models ended on October 23, 2026. The affected IDs span GPT-3.5, GPT-4, GPT-4.1 nano, GPT-4o, GPT Image, o1, o3-mini and o4-mini.
- Fine-tuned variants of affected base models follow the same shutdown schedule; OpenAI's SFT/DPO guides name gpt-4.1-nano, and its RFT guide names o4-mini.
- The image-prompting guide and Images API reference now both state that gpt-image-1 shut down on October 23; the reference labels it deprecated.
- The public docs do not specify an exact cutoff time, time zone or account-specific error response. Dedicated capacity may be available in some cases, but must be confirmed separately.
- Inventory the model IDs your requests actually send, check your authorized project and validate a replacement against real tasks. Changing network egress will not restore a retired model.
Which model IDs shut down on October 23?
OpenAI's current deprecations page says API access to the legacy models announced on April 22, 2026, shut down on October 23. The listed IDs include gpt-3.5-turbo-0125, gpt-4-0613, gpt-4-1106-preview, gpt-4-turbo, gpt-4.1-nano and gpt-4.1-nano-2025-04-14, gpt-4o-2024-05-13, gpt-image-1, o1-2024-12-17, o1-pro-2025-03-19, o3-mini-2025-01-31, o4-mini-2025-04-16 and ft-o4-mini-2025-04-16.
The deprecations table gives model-specific replacement suggestions. For example, it points from gpt-4.1-nano to gpt-5.6-luna, from o4-mini-2025-04-16 to gpt-5.6-terra, and from gpt-image-1 to gpt-image-2.5-sunburst or gpt-image-2.5-flare. A recommendation is a migration direction, not a promise of matching capabilities, tool behavior, prices or quotas.
Fine-tuned models follow their base model; training has a separate cutoff
OpenAI's deprecations page says fine-tuned variants of affected base models follow the same shutdown schedule. The SFT and DPO guides name gpt-4.1-nano-2025-04-14 and its fine-tuned variants; the RFT guide names o4-mini-2025-04-16 and its fine-tuned variants. Inference on a fine-tuned model ends when its base model shuts down.
Training eligibility and inference availability are separate. OpenAI's current documentation sets January 6, 2027, as the end date for eligible existing customers to create new fine-tuning jobs, but that does not extend inference on a base model that shut down on October 23. New organizations and organizations without recent fine-tuned inference also face separate eligibility limits; check the current rules before planning a job.
The image docs now confirm that gpt-image-1 is shut down
OpenAI's image-prompting guide now labels gpt-image-1 retired and says it shut down on October 23, 2026. The Images API reference states the same status and recommends gpt-image-2.5-sunburst or gpt-image-2.5-flare.
The API reference retains gpt-image-1 as a deprecated enum value; that does not put it in the current supported-model list. Follow the current support notes and your project permissions instead of treating an older SDK enum as proof that a model can still be called.
A shutdown date is not an account-specific response receipt
OpenAI's public deprecations page confirms the platform-level shutdown date, but it gives no exact time or time zone for October 23 and does not say which HTTP status or error text a particular account will receive. OpenAI also says developers may be able to provision dedicated capacity for continued access after a shutdown date in some cases. That requires separate confirmation; it is not an automatic extension for every organization.
If your organization expects dedicated capacity or another arrangement, confirm the exact model, project and validity period through an approved OpenAI contact channel. The public deprecations page does not prove that your project has an exception and cannot replace its actual response records.
Inventory the model IDs your application really sends
Check the model field in request bodies, configuration, scheduled jobs, batch tasks, fallback chains and saved fine-tuned model IDs. Searching only a display name can miss versioned snapshots, background work or custom IDs beginning with ft:.
The Models API reference defines shutdown_date as the model's shutdown date or null. This documented lifecycle field is not proof that a specific project can make the call. An authorized project can check its current model record and actual response; do not generalize one project result to every organization or paste an API key into a ticket or public log.
Validate replacements against real tasks and separate network errors
Check OpenAI's suggested replacement for each ID, then use low-risk samples to compare output quality, tool calls, structured results, context needs, cost and rate limits. Switch gradually after the behavior meets your acceptance criteria. Keep a change record and a fallback model verified as available to the current project; do not rely on a retired ID as a working fallback or replace every ID just because the names look similar.
When troubleshooting, record the time, exact model ID, HTTP status, response body and request ID. Then check permissions, model eligibility, quota and OpenAI status. A retired model, missing project access or an incompatible parameter is not a network fault. Investigate DNS, TLS, timeouts or proxy errors only when you have connection-layer evidence; changing network egress will not restore a retired model.
Sources
Frequently Asked Questions
Does upgrading the SDK complete the model migration?
An SDK version alone does not establish that. Check the model field your requests send and the fixed IDs in scheduled jobs, batches and fallback configurations, then validate the replacement against the original task.
Can I rerun an entire batch that failed because of an old model?
First determine which original tasks completed, which definitively failed and which still have an unknown outcome. Resume only the work confirmed to need it, taking particular care to avoid repeating external actions. Migrating the model does not prove that the previous batch produced no results.
What should I share when asking a community or support team for help?
Provide a minimal, redacted description, time and time zone, exact model ID, HTTP status and request ID. Remove API keys, authorization headers, private prompts and business data; do not upload complete raw logs.