GPT-4o API Shutdown: What Small Businesses Must Do
What is happening? OpenAI will remove GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini from its API on 14 January 2027. Any chatbot, Google Sheet or automation that calls these models by name will stop working. The fix is to change the model name, usually to gpt-6-luna, and test before the deadline.
This GPT-4o API shutdown does not touch the ChatGPT app. It only hits tools that talk to OpenAI behind the scenes. If you run a small business with a chatbot, a lead sorter or an AI step in Zapier, Make or n8n, this post is for you.
I build CRMs and Google Sheets automations for small businesses, so I see how often a model name is typed once and forgotten. A deadline like this is exactly when those forgotten lines break. Below is a plain plan you can follow in an afternoon.
What is the GPT-4o API shutdown?
An API is a door that lets one program talk to another. When your chatbot asks OpenAI for an answer, it goes through the API and names the model it wants, such as "gpt-4o". On 14 January 2027 that door closes for four models. Calls that name them will fail.
OpenAI sent this notice to developers on 14 October 2026. You can read the list on the OpenAI deprecations page. Here is what it says, in my own words.
| Model | Stops working | OpenAI suggests |
|---|---|---|
| GPT-4o | 14 January 2027 | gpt-6-luna |
| GPT-4o mini | 14 January 2027 | gpt-6-luna |
| GPT-4.1 | 14 January 2027 | gpt-6-luna |
| GPT-4.1 mini | 14 January 2027 | gpt-6-luna |
| GPT-4o, version dated 2024-05-13 | 23 October 2026 (today) | gpt-5.6-sol |
The last row matters if your tool names that exact old version. It stops first. Most people use the newer versions, but check.
Does the shutdown affect me if I only use ChatGPT?
No. This is an API change. Datanorth reports that OpenAI already removed these models from the ChatGPT app earlier in 2026, so app users saw that change months ago.
You are affected if any of these is true:
You have a website chatbot or a WhatsApp bot that uses OpenAI.
You have a Google Sheet or Apps Script that sends text to OpenAI.
You use Zapier, Make, n8n or a similar tool with an OpenAI step.
You paid a developer or agency to build an AI feature and you do not know which model it uses.
That last point is the risky one. If someone else built it, ask them for the model name in writing.
How do you find where your business uses GPT-4o?
The model name is usually a short piece of text. It can hide in a few places. Work through this list in order.
Your OpenAI account. Open the usage page and look at spending by model. If gpt-4o or gpt-4.1 shows up, something is calling it.
Your automation tools. Open each workflow that has an OpenAI or ChatGPT step. The model is normally a drop-down or a text box.
Your scripts. In an Apps Script or any code file, search for the words gpt-4o and gpt-4.1. In my experience with Apps Script, the model name is often one text value near the top.
Your chatbot platform. Some platforms hide the model. Ask the vendor which model they use and when they will move.
Scheduled jobs. Anything that runs at night or once a month is easy to forget. List them.
If you want a quick place to search a long script, paste it into the FRQ Web IDE and use the find box. Here is the whole process on one page.
Five step plan to move off GPT-4o before 14 January 20271. FindList every tool that names gpt-4o or gpt-4.12. AskGet model names from vendors and developers3. TestRun real examples through the new model4. SwitchChange the model name, one workflow at a time5. WatchCheck results and cost for two weeksFRQ Tools. Finish step 4 well before 14 January 2027.Do the steps in order. Keep the old name written down until the new one works.What should you switch to?
OpenAI recommends one replacement for all four models: gpt-6-luna. Its model page calls it the most efficient model, built for focused, high volume tasks. That fits most small business jobs, such as sorting leads, writing short replies and tagging messages.
Luna is also much cheaper on paper. These are the input prices per one million tokens. A token is a small piece of text, roughly three quarters of a word. The numbers come from the Digital Applied price table, which it read from OpenAI on 15 October 2026. Prices can change, so check OpenAI's own pricing page before you plan a budget.
Bar chart: input price per one million tokens for old models and gpt-6-luna, in US dollarsInput price per 1 million tokens (US dollars)GPT-4o$2.50GPT-4.1$2.004.1 mini$0.404o mini$0.15Luna$0.10Source: Digital Applied, citing OpenAI list prices read 15 Oct 2026. Bars to scale.Input price only. Output prices and extra thinking tokens also count, so test on your own work.Do not expect a bill that is exactly lower. Luna has a setting called reasoning effort, which controls how much the model thinks before it answers. OpenAI's model page says it can be set from none up to max. More thinking can use more tokens. Digital Applied also notes that these thinking tokens are billed as output. So run a small test and look at the real cost.
How do you test before you switch?
Never change the model on a live customer workflow without a test. A new model can answer in a different style, a different length or a different format. A short test catches this.
Collect real examples. Pull twenty to fifty real inputs from your chat logs, your sheet or your inbox. Include the messy ones.
Run them through both models. Put the old and new answers side by side.
Check what your system needs. If another step reads the answer, such as a CRM field or a spreadsheet column, check that the format still matches.
Compare length. Paste long answers into the Word Counter to see if replies got longer or shorter. Long replies can break a WhatsApp or SMS limit.
Clean your sample text. Copied text often has odd spaces and line breaks. The Text Formatter and Cleaner tidies them so your test is fair.
If your tool uses function calling, where the model fills in fields for another program, test that with extra care. Digital Applied reports that on the older Chat Completions endpoint, function calling works only when reasoning is set to none. Read the notes for the endpoint you use.
What happens to fine-tuned models?
A fine-tuned model is a copy of a model that you trained on your own examples. OpenAI's notice says these follow the same schedule as the base model. So a fine-tune built on GPT-4o mini stops on 14 January 2027 too.
This is the hardest case. Datanorth reports that gpt-6-luna cannot be fine-tuned, that OpenAI is winding down its self-serve fine-tuning, and that new training jobs can start only until 6 January 2027. If you have a fine-tune, decide now. A common option is to move the examples into the prompt, so the model sees a few good samples every time. Another is to search your own documents and pass the right part to the model.
For many small businesses a clear prompt with three good examples is enough. Test it before you give up on the fine-tune.
Should you wait until January?
No. A dead model does not give a polite warning. A job that runs at night on 15 January will simply fail, and nobody will see it until a customer complains. Chatbots that go quiet are a lead leak. I wrote about this in WhatsApp lead management for small business: the chat is not the record, so you need a way to know when a message got no reply.
Here is a simple timetable that leaves time for mistakes.
| When | What to do |
|---|---|
| This week | Find every use of gpt-4o and gpt-4.1. Ask vendors and developers. |
| By mid November | Run your test examples. Decide on fine-tunes. |
| By mid December | Switch one workflow at a time. Keep the old name written down. |
| Before 6 January | Last day to start a new fine-tune job, if you still need one. |
| 14 January 2027 | The four models stop working. |
What if a tool you pay for uses the old model?
Ask the vendor three plain questions. Which model do you use for my account? When will you move to a new one? Will my replies or costs change? Put the answers in an email so you have a record. If a vendor cannot answer, treat that as a risk and plan a backup.
If you built your own chatbot, my guide on launching a chatbot in five minutes covers the platform side. If you use WhatsApp for support, the WhatsApp service message charges post explains that bot cost is a second bill to watch next to the OpenAI one.
The same rule helps for spreadsheets. If you run a lead sheet with scripts, add a "model" cell at the top and read the name from it. Then one change fixes every script. My post on when a Google Sheets CRM wins shows where these scripts fit in a small business.
The short version
Four OpenAI models leave the API on 14 January 2027. ChatGPT users are not affected. Find where your tools name the old models, test with real examples, switch to gpt-6-luna one workflow at a time, and handle fine-tunes early. Start this week, not in January.
If you would rather not do this alone, I can audit your automations and chatbots, move them over and add alerts so a failure never goes unnoticed. Get in touch about custom CRM and business automation work and tell me what you run today.
Related reading
Want this built for your business?
Tell us what slows you down. We reply on WhatsApp within an hour with a plan and a fixed price.
