OpenAI Is Retiring Five GPT-4 API Snapshots on January 14, 2027

Source announcement: 14 October 2026 | Checked: 31 October 2026 | Scheduled shutdown: 14 January 2027 | Status: Deprecated, still accessible until the stated shutdown date

Here is a deadline that deserves a place on every AI application's release calendar. OpenAI has scheduled selected GPT-4o and GPT-4.1 API model snapshots for shutdown on 14 January 2027. The announcement was published on 14 October, not today. But the operational question is urgent now: which production workflows, integrations, and fine-tuned models still depend on those identifiers?

At the time of this review on 31 October, the scheduled shutdown is 75 calendar days away. That is not much time when a model powers customer-facing chat, document extraction, HR workflows, structured outputs, or automated decisions. The risk is not that every ChatGPT user suddenly loses GPT-4o; this notice concerns specified API models and their dependent applications.

Exactly which API models are affected?

OpenAI's official API deprecations register identifies these model snapshots and recommends gpt-6-luna as the replacement:

Model family	Listed model snapshot(s)	Shutdown date
GPT-4o	gpt-4o-2024-08-06, gpt-4o-2024-11-20	14 Jan 2027
GPT-4o mini	gpt-4o-mini-2024-07-18	14 Jan 2027
GPT-4.1	gpt-4.1-2025-04-14	14 Jan 2027
GPT-4.1 mini	gpt-4.1-mini-2025-04-14	14 Jan 2027

Important distinction: Do not assume every similarly named model, endpoint, or ChatGPT experience is covered by this exact table. Inventory the concrete model IDs your applications actually call, and use the current model catalog to check candidate replacements. The published deprecation register, not a social-media summary, is the authority for dates and identifiers.

Fine-tuned GPT-4 models need special attention

OpenAI states that fine-tuned versions of the affected base models follow the same 14 January shutdown schedule. This can be easy to miss: a team may see a custom fine-tuned model ID in its configuration and forget which underlying base model it uses.

OpenAI's fine-tuning documentation also says the self-serve fine-tuning platform is being wound down. Eligible existing customers can create new fine-tuning jobs only until 6 January 2027, unless the base model shuts down sooner. Existing fine-tuned inference is supported only until its base model's retirement. Treat the fine-tuned model migration as a separate workstream; swapping a model ID does not recreate learned behavior or guarantee identical outputs.

What happens if you do nothing?

At the scheduled shutdown, the listed API models will no longer be accessible, according to OpenAI's published definition of shutdown. Applications that continue to request retired IDs risk request failures and service interruption. The precise error handling, fallback behavior, and customer impact depend on each application's implementation, so test those behaviors rather than assuming an automatic substitution.

This is particularly consequential when a model is embedded in a workflow rather than used directly by a person. A hiring assistant, document classifier, customer-service integration, or scheduled agent may continue to run unattended until its next request fails. The business owner may not even know that an old model name is buried in an environment variable, deployment secret, workflow configuration, or third-party connector.

A practical migration plan before 14 January

1. Inventory actual usage now. Search application configuration, code, infrastructure variables, workflow tools, SDK defaults, fine-tuned model IDs, and third-party integrations for the listed snapshots. Cross-check against actual API usage and model inventory rather than relying only on code search. OpenAI documents an optional shutdown_date field on model records in its deprecation guidance.
2. Classify business criticality. Identify which use cases would fail visibly, which would silently degrade, and which need regulated or contractual review. Assign an owner and migration deadline to each dependency.
3. Test the recommended replacement, do not assume parity. OpenAI recommends gpt-6-luna for the listed snapshots, but output style, reasoning, token usage, latency, tool interactions, and structured-output behavior still need application-specific validation. Follow OpenAI's model selection guidance and compare on representative inputs.
4. Build a regression set. Test difficult real-world prompts, multilingual content, malformed documents, safety refusals, JSON schemas, function calls, and any human-approval steps. Measure quality and failure rates, not just whether the new model responds. Use the evaluation best-practices guidance to design reproducible checks.
5. Run a controlled cutover. Put the replacement behind a configuration flag or staged rollout, monitor errors and cost, and retain a rollback plan that does not depend on the retired model after 14 January. Review OpenAI's production guidance before deployment.
6. Handle fine-tuned models explicitly. Preserve datasets, task definitions, evaluation cases, and the reasons the original model was fine-tuned. Decide whether prompt engineering, retrieval, or another supported model can meet the acceptance criteria before the underlying base model expires.
7. Set an internal deadline earlier than the vendor deadline. Finish production migration with enough buffer to observe real traffic, handle vendor or customer sign-off, and correct regressions before 14 January.

Does this mean you must migrate from Chat Completions to Responses API?

Not necessarily as part of this specific model-retirement notice. A model retirement and an API-interface migration are separate decisions. The immediate requirement is to remove dependencies on the listed model IDs. Teams considering broader modernization can review OpenAI's Responses API migration guidance, but should scope that work independently so an endpoint rewrite does not obscure a time-critical model change.

Availability, commercial terms, and exceptions

The replacement is recommended in OpenAI's documentation; that does not guarantee identical features, pricing, rate limits, or availability for every organization. Check the model catalog, pricing, and your own API account before committing to a target architecture. OpenAI also notes that dedicated capacity may sometimes be possible after a shutdown date through sales; this is not a general extension or an entitlement.

My practitioner takeaway

I would not treat this as a routine model upgrade ticket. I would treat it as a dependency-removal exercise with a named business owner, a verified model inventory, and measurable acceptance tests. The best time to discover an old GPT-4o snapshot is during an October inventory, not when an automated workflow stops responding in January.

Bottom line: The deprecation was announced on 14 October 2026; the listed API models and their fine-tuned derivatives are scheduled to shut down on 14 January 2027. The action today is to identify dependencies, test replacements, and complete a controlled migration well before the deadline. For any subsequent date or model-list changes, recheck the live OpenAI deprecations register.