Every language model version in the API has a date after which requests to it stop working: on October 14, 2026, OpenAI announced that GPT-4o, GPT-4o mini, GPT-4.1, and GPT-4.1 mini will be removed from the API on January 14, 2027, and more than a dozen older models will be turned off on October 23, 2026. To make sure a model switch doesn’t turn into an outage, you need to know in advance where model IDs are hardcoded in your systems, keep them in a registry with deprecation dates, maintain a fallback model, and run a benchmark set of tasks on the old and new versions before switching over. According to a study of 17,703 public repositories, 82% of migrations happen only after shutdown—when the application is already broken.
Prepared by AIrassvet October 15, 2026 based on the official model shutdown pages from OpenAI, Anthropic, and Google, the preprint “When the Model Retires” (arXiv, September 25, 2026), and a Habr review dated October 15, 2026. It is intended for IT leaders and product owners whose chatbots, AI agents, RAG assistants, or integrations run on language model APIs. The dates in the article were verified against the providers’ pages as of publication; they change, so check the original source before planning. We did not carry out our own migrations for this article: the preparation plan is the editorial team’s methodology. AIrassvet is not affiliated with the providers or authors mentioned; the section at the end of the article describes the publisher’s services.
Contents
- In brief
- What’s being retired in the coming months
- How providers manage a model’s lifecycle
- How teams actually migrate
- What breaks when you change models
- Model switch preparation plan
- Checklist for January 14, 2027
- FAQ
- Primary sources and verification method
In brief
- GPT-4o and GPT-4.1 are leaving the API on January 14, 2027. OpenAI announced this on October 14, 2026, and recommends gpt-6-luna as the replacement; fine-tuned versions of these models will be retired on the same schedule.
- Another group of OpenAI models is being shut down on October 23, 2026, including gpt-4-turbo, gpt-4.1-nano, o1, o3-mini, and o4-mini.
- Providers give different notice periods: for OpenAI public models it is at least 6 months, for Anthropic at least 60 days, and Google’s table lists the earliest possible dates.
- Most teams respond too late. In a study of 22,555 commits, 82% of migrations away from retired models were made after the shutdown date, and in 94% of migrating applications the model ID was hardcoded directly in the code.
- The most dangerous failure is the silent one. The new model responds without errors, but differently: formatting, length, and refusal rates change. The only way to catch that is with a benchmark task set.
What’s being retired in the coming months
Below are the dates from the providers’ official pages as of October 15, 2026 (OpenAI Deprecations, Anthropic Model deprecations).
| Date | Provider | What’s being retired | Recommended replacement |
|---|---|---|---|
| October 23, 2026 | OpenAI | gpt-3.5-turbo, gpt-4, gpt-4-turbo, gpt-4-1106-preview, gpt-4.1-nano, gpt-4o-2024-05-13, o1, o1-pro, o3-mini, o4-mini, gpt-image-1, and a number of fine-tuned models | gpt-5.6-sol, gpt-5.6-terra, or gpt-5.6-luna depending on the model |
| November 30, 2026 | Anthropic | claude-sonnet-4-5-20250929 | claude-sonnet-5-5 |
| January 14, 2027 | OpenAI | gpt-4o (snapshots 2024-08-06 and 2024-11-20), gpt-4o-mini-2024-07-18, gpt-4.1-2025-04-14, gpt-4.1-mini-2025-04-14, and their fine-tuned versions | gpt-6-luna |
| April 1, 2027 | OpenAI | gpt-5.3-codex, gpt-5.4-nano, gpt-5.1 | gpt-6-sol or gpt-6-luna |
In addition to the announced retirements, Anthropic lists the earliest possible retirement date for each active model — “not earlier than.” For example, for claude-opus-4-5-20251101 it is November 24, 2026, and for claude-haiku-4-5-20251001 it is October 15, 2026. This is not an announced shutdown, but it is a signal: models like these should already be on your migration plan.
A separate note on intermediaries. Anthropic notes that the dates on its page apply to its own platforms, while Amazon Bedrock and Google Cloud use their own retirement schedules. If you access a model through a cloud provider or aggregator, follow its calendar.
How providers manage a model’s lifecycle
Deprecation means a model has been announced for retirement: it still works, but a replacement and a date have already been announced. Retirement or shutdown is the point after which requests to the model return an error. That is the migration window.
| OpenAI | Anthropic | Google Gemini API | |
|---|---|---|---|
| Minimum notice | Public models — at least 6 months; specialized variants — at least 3 months; preview — may be around 2 weeks | At least 60 days for publicly released models | The table shows the earliest possible dates; the exact date is announced in advance |
| Where it is announced | Deprecations page and emails to active users | Model deprecations page and emails | Release notes page; dates are on the Deprecations page |
| What else matters | The legacy status means the model is no longer updated and will be retired later | Dates for Bedrock and Google Cloud may differ | The date in the table may shift: it is the earliest possible deadline |
Sources: OpenAI, Anthropic, Google. OpenAI also states that in the event of security threats or regulatory requirements, a model may be retired sooner, and it does not recommend preview models for business-critical tasks unless you are ready to migrate on short notice.
A Habr review also cites the actual median notification windows based on independent calculations by Okane Land: 182 days for OpenAI, 63 days for Anthropic, and 28 days for Google (Habr, 10/15/2026). We did not verify the original source of these calculations, so use them as a reference, not a guarantee.
How teams actually migrate
The preprint “When the Model Retires: An Empirical Study of LLM Migration in Open-Source Applications” (Hyungjin Lukas Kim, arXiv, September 25, 2026) is the first major study of what happens to applications when a model is shut off (arXiv:2609.31288). The author found GitHub commits that move code away from officially retired OpenAI, Anthropic, and Google models and matched them with announcement and shutdown dates.
| What was measured | Result |
|---|---|
| Data volume | 22,555 commits across 17,703 repositories in 2024–2026; 5,139 matched to official events |
| Migrations after the shutdown date | 82% (95% confidence interval 79–84%) |
| Share of late migrations with Anthropic notice of 60–114 days | 89% |
| Share of late migrations with OpenAI’s one-year notice for Assistants API | 13% |
| Applications with the model identifier directly in code | 94% |
| Size of changes during migration (median) | from 6 added lines for “prompt-only” apps to nearly 700 for fine-tuned models |
| Migrations that changed providers | 8% |
Two more takeaways from the abstract: the share of late migrations does not depend on repository popularity, prior shutdown experience, or even whether there is an abstraction layer over the provider, and each increase in the notification period by about 2.7x reduces the odds of a late migration by roughly three-quarters. This is a single-author preprint without peer review, so the numbers should be read as a strong signal, not a rule.
Main practical takeaway: long notice helps, but most teams still do not start working in advance. That means the shutdown date needs to become a task on the team calendar on the day the model is connected, not the day the email arrives.
What breaks when you switch models
The Habr review identifies four failure scenarios (Habr, 10/15/2026); we added what can be seen in official announcements.
- Hard shutdown. A request to the retired model returns an error — Anthropic explicitly says requests to retired models fail. The user sees a broken feature.
- Behavioral drift after replacement. The new model technically works, but responds differently: different JSON format, length, style, or refusal rate. There are no errors, but process quality drops.
- Price changes. The replacement may cost differently, and the budget has to be recalculated with each transition.
- Retiring an entire API. OpenAI shut down the Assistants API on August 26, 2026, after giving one year’s notice: this is not a config change, but a rewrite of the integration for the Responses API and Conversations API.
The more infrastructure wrapped around a model, the more expensive the transition: according to the preprint, an app with prompts only usually needs just a few lines changed, while a system with a fine-tuned model may need hundreds. For fine-tuned OpenAI models, a separate rule applies: they are retired together with the base model, so fine-tuning will have to be repeated on the new base.
Model change preparation plan
The plan below is the AIrassvet editorial team’s methodology, based on provider recommendations and the preprint’s findings. It is not the result of our own project.
- Inventory. Find every place where models are called: code, configurations, low-code workflows, prompts in databases, agent settings, and third-party services. Mark fine-tuned models and models accessed through intermediaries separately.
- Model register with dates. For each identifier, record the provider, purpose, owner, the “not before” or announced retirement date, and the recommended replacement. The only source for dates should be official provider pages.
- Put the identifier into configuration. Move model names out of code and into settings so switching models becomes a configuration change, not a release. Version-pinned models with dates behave predictably; aliases like “latest” change without your involvement.
- Fallback model. For critical scenarios, choose and test a second model in advance that you can switch to if the primary one fails.
- Reference task set. Assemble a set of real prompts with verifiable criteria: response format, required fields, correct facts, acceptable refusal rate. Run it on both the old and new model before switching. We covered how to build such a set in the article about our own benchmark for choosing an LLM.
- Announcement monitoring. Check provider Deprecations pages and emails to the account address every week; when a model in the register appears, create a task with an owner and deadline.
- Budget recalculation. Compare the new model’s price with token usage on the reference set; we wrote about forecasting costs as token usage grows in the article about stable AI costs.
If the company uses multiple providers, some of these steps are easier to cover with a single routing layer; how it works and what it does not solve are covered in the article about model routers. Remember that such a layer does not prevent provider shutdowns and does not guarantee identical model behavior.
Checklist for January 14, 2027
For systems on GPT-4o, GPT-4o mini, GPT-4.1, and GPT-4.1 mini — the minimum action plan compiled by the editorial team:
| Step | Question | Done |
|---|---|---|
| 1 | Have all calls to gpt-4o, gpt-4o-mini, gpt-4.1, and gpt-4.1-mini, including fine-tuned versions, been found? | |
| 2 | Does every call have an owner and a scenario criticality level? | |
| 3 | Has a reference task set been built for each critical scenario? | |
| 4 | Has the set been run on the recommended replacement, and have format, quality, latency, and price been compared? | |
| 5 | Does the fine-tuning need to be repeated on the new base model? | |
| 6 | Has the model identifier been moved into configuration, and is there a backup model? | |
| 7 | Has the switchover date been set well before January 14, 2027, with time to roll back? |
If your systems still have models from the October 23, 2026 list, start with those: there is just over a week left before shutdown.
Frequently Asked Questions
When will OpenAI shut off GPT-4o in the API?
According to OpenAI’s Deprecations page, the gpt-4o-2024-08-06 and gpt-4o-2024-11-20 snapshots, as well as gpt-4o-mini-2024-07-18, gpt-4.1-2025-04-14, and gpt-4.1-mini-2025-04-14, will be removed from the API on January 14, 2027; the recommended replacement is gpt-6-luna. The gpt-4o-2024-05-13 snapshot is being turned off earlier, on October 23, 2026.
What happens to requests sent to a deprecated model?
They will fail with an error: there is no automatic fallback to a new model. That is why you need a date registry, a backup model, and error monitoring so you can catch the problem before users do.
How much notice is given before a model is shut down?
Stated minimums: OpenAI gives at least 6 months for generally available models, 3 months for specialized variants, and about 2 weeks for preview models; Anthropic gives at least 60 days. Google lists the earliest possible dates and announces the exact date in advance.
Can you just replace the model name in code?
Technically, often yes, but the new model may respond differently: different format, length, or refusal rate. Before switching, run a benchmark suite on the old and new models and compare the results, otherwise you may have a silent failure with no errors.
What should be done with fine-tuned models?
At OpenAI, fine-tuned versions are deprecated together with the base model. You will have to retrain on the new base or replace the approach with one that does not rely on fine-tuning, so these systems should be included in the migration plan first.
Can a model aggregator help you avoid migration?
No. An aggregator or router makes it easier to switch between models, but it does not override provider deprecations and does not guarantee identical behavior. A benchmark suite and a migration plan are still required.
Primary sources and verification method
Method: on October 15, 2026, we read the OpenAI, Anthropic, and Google model deprecation pages, the arXiv preprint abstract, and a Habr article, extracted dates, notification policies, and research figures, and cross-checked the dates from the article against official pages. Where the article’s data were not confirmed by the primary source, that is noted in the text.
- OpenAI: Deprecations — notice periods, announcements from October 14, 2026, October 1, 2026, and April 22, 2026, Assistants API shutdown.
- Anthropic: Model deprecations — model status, notice period, Claude Sonnet 4.5 schedule, and the note about partner platforms.
- Google: Gemini deprecations — announcement sequence and the meaning of the dates.
- arXiv, 09/25/2026: Kim, "When the Model Retires: An Empirical Study of LLM Migration in Open-Source Applications".
- Habr, 10/15/2026: "The Threat to AI Agents: LLMs Are Being Taken Out of Production After Just 6 Months" — failure scenarios, Okane Land medians (published in the vendor’s corporate blog for a unified API).
Limitations: provider schedules change; the preprint has not been peer reviewed; the notice-window medians come from a secondary source.
How AIrassvet helps prepare AI systems for a model change
Migration goes smoothly when you already know where the models are running, how to validate the replacement, and where to switch if something fails. On the AIrassvet website the following areas are described that relate directly to this task:
- process and integration audits: inventory of model calls, owners, and scenario criticality;
- integration and testing: moving AI agents and RAG systems to a new model with validation against a benchmark suite before the switch;
- deployment and fine-tuning of language models, including on the customer’s own servers, for systems that need to be moved off a fine-tuned model or made independent of an external API;
- post-launch support.
The first step is to choose one process that runs on a model from the deprecation list, document its current quality and data sources, define the constraints, and agree on what result from the new model will be considered acceptable.
Bottom line: the shutdown date is part of the architecture
A model in the API is a dependency with an expiration date. On January 14, 2027, GPT-4o and GPT-4.1 will leave the OpenAI API, and a study of thousands of repositories shows that most teams migrate only after something breaks.
You can protect yourself without a scramble: a model registry with dates, identifiers in configuration, a backup model, a benchmark suite, and weekly checks of provider pages. Then model changes become a planned task, not an incident.
