Will your AI integrations stop on 14 January? What to do now

In short

• OpenAI is removing GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini from its API on 14 January 2027. Software and automations that call these models will stop working on that date.
• If your team only uses ChatGPT in the app or browser, nothing changes: these models already left ChatGPT in February 2026.
• Spend an hour finding out which integrations are affected, give each one an owner, and test the replacement, GPT-6 Luna, on your own examples well before the Christmas break.

On 14 October 2026, OpenAI announced that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini will be removed from the API on 14 January 2027 (OpenAI, Deprecations). The API is how software talks to OpenAI's models. If you run an automation in n8n, Make or Zapier, a chatbot on your website or your own software built on one of these models, it will stop on that day unless someone switches the model. OpenAI recommends GPT-6 Luna as the replacement.

The switch itself is usually small: change a model name and test. The real risk is what you don't know. Which integrations are running, who looks after them, and will the new output be just as good? If something still runs on an older model, such as gpt-3.5-turbo, gpt-4 or o4-mini, it is more urgent: those stop on 23 October 2026.

Who is affected, and who isn't?

Anything that calls one of these models with an OpenAI API key is affected:

• automations in n8n, Make or Zapier with an OpenAI step
• a chatbot or form on your website that writes answers or summaries
• your own software or customer portal with a GPT integration
• fine-tuned models: versions of these models trained on your own data. They shut down on the same date.

People who use ChatGPT itself, in the app or browser, are not affected. OpenAI removed GPT-4o and GPT-4.1 from ChatGPT on 13 February 2026, saying at the time that there were "no changes at this time" in the API (OpenAI, Retiring GPT-4o and older models). This announcement covers the API only. GPTs you built inside ChatGPT are also outside its scope.

If your integration runs through Microsoft Azure rather than directly through OpenAI, Microsoft's own schedule applies. Check that separately.

How can you check in an hour whether you're affected?

1. List your API keys. On the OpenAI platform, see which keys exist and in which project. Note which tool or application uses each one. A key nobody can explain is your first action point.
2. Check usage by model. The usage dashboard shows consumption per project. A developer can break it down by model and by key through the Usage API. If gpt-4o, gpt-4o-mini, gpt-4.1 or gpt-4.1-mini appears, you are affected.
3. Open the settings in your automation tools. In n8n, Make and Zapier the model is set per step. Go through every workflow with an OpenAI step and note the model. One catch: OpenAI lists only dated versions, such as gpt-4o-2024-08-06. Many tools use the short name gpt-4o, which points to one of those versions. OpenAI doesn't say what happens to the short names, so assume they stop too.
4. Have the code searched. A developer searches code and configuration, including environment variables, for "gpt-4o" and "gpt-4.1".
5. Ask your suppliers. If an agency or vendor built something with AI for you, ask three questions: which model does it run on, when will you switch, and how will you test it? The same goes for third-party chatbots on your site.

Put the results in one list: integration, model, owner, and how much the business depends on it.

What other deadlines are coming?

The same OpenAI page lists more dates. The first is as early as 23 October 2026:

The text-to-speech models need more than a name change, because the replacement runs on the Realtime API, a different interface. If you use speech for a phone assistant or voice-overs, treat it as a small project. Developers can also monitor these dates automatically: GET /v1/models returns a shutdown_date for each model once OpenAI has announced one.

How do you switch without surprises?

1. Give each integration an owner. One person tests, approves the switch and keeps an eye on it afterwards. If a supplier does the work, this person chases their answer.
2. Collect your own test cases. Take twenty to fifty real examples from recent months, with the output that was right. Include the awkward ones.
3. Run them on GPT-6 Luna and compare. Put old and new output side by side and check content, tone, length and format. If a later step reads the output automatically, for instance as JSON, test that step too.
4. Adjust prompts and settings. GPT-6 Luna is a reasoning model that thinks at "medium" effort by default (OpenAI, GPT-6 Luna). For simple work such as classification, a lower setting may be enough, and faster. Test it. Watch out with older integrations: in Chat Completions, function calling (the model asking your software to take an action) only works with effort set to "none". The Responses API has no such limit.
5. Check the costs. GPT-6 Luna costs 0.10 dollars per million input tokens and 0.50 dollars per million output tokens, half that through Batch (OpenAI, API Pricing). Requests over 272,000 input tokens cost more. The model's reasoning also counts towards usage, so look at actual consumption after a week of testing. If you process data in Europe through data residency, check the 10% uplift on newer models. GPT-6 Luna supports EU processing, subject to OpenAI's approval.
6. Reconsider fine-tuned models. GPT-6 Luna cannot be fine-tuned, and from 6 January 2027 you can no longer start new fine-tuning jobs. A GPT-4o fine-tune therefore stops with no direct successor. Test whether a clear instruction with examples in the prompt gives the same result on your test set.
7. Switch, then watch closely for a week. Plan the change well before 14 January, ideally before Christmas, so there is time to adjust. In the first week, spot-check the output and the error logs.

How GPT-6 Luna compares with Sol and Astra, and what you can use in Europe, is covered in GPT-6.1 Sol, Ultrafast and the Agents API: what to choose.

When is this not a problem?

• You only use ChatGPT, Copilot or another AI app, with no integration of your own.
• Your integrations already run on a GPT-5.6 or GPT-6 model. Do check the other dates in the table.
• Your supplier has confirmed in writing that they will switch and test before the deadline.
• It's an old experiment nobody uses any more. Revoke the API key instead.

The real risk is an integration nobody knows exists. You only find out on 14 January, when a customer gets no reply or a process quietly stops.

Hello Growth builds and maintains integrations and automations. Want to go through your integrations together or have the switch tested? Book a call.

Sources

• OpenAI, Deprecations, announced 14 October 2026
• OpenAI, Retiring GPT-4o and older models
• OpenAI, GPT-6 Luna
• OpenAI, API Pricing
• OpenAI, Your data
• OpenAI, Reviewing API usage and costs
• OpenAI, Usage API: completions