GPT-4o retirement: what to check if your business uses it

What OpenAI has announced

On 14 October, OpenAI told developers using those models that they will be removed from its API, the doorway other software uses to plug into its artificial intelligence. Its official model deprecations page lists the affected versions and the shutdown date.

• Models affected: GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini, across five dated versions, according to OpenAI.
• Shutdown date: 14 January 2027. Fine-tuned versions of those models, the ones a company has trained on its own examples, follow the same schedule.
• Recommended replacement: GPT-6 Luna for all four models. According to DataNorth, GPT-6 Luna cannot be fine-tuned.
• Other retirements come first: under an April notice, gpt-4.1-nano and the May 2024 version of GPT-4o leave the API on 23 October 2026, according to the same OpenAI page.

The notice only covers the API. According to DataNorth, people who only use ChatGPT do not need to change anything because of it. For those who do switch models, Superpower Daily advises testing the replacement on representative requests and edge cases before making the move.

The order tool nobody quite understands

Imagine you run a family wine distributor with four people. Eighteen months ago, a freelance developer set up a small automation for you: it reads the order emails from bars and restaurants, works out which wines and how many cases they want, and adds them to the order sheet on its own. Since then, Elena, who runs the office, no longer copies every Monday morning order by hand.

It works so well that nobody thinks about it. The developer has moved on to another company. The OpenAI account sits under an email address nobody checks. And if someone asked Elena which AI model her automation uses, she would have no idea. That is not carelessness on her part: she has simply never needed to know.

Why do so many automations hang on a single model?

The scene is made up, but it looks a lot like many we come across. When someone connects a piece of software to OpenAI's AI, they write the exact name of the model into the code. During testing they pick whichever responds best that month, everything works and the project is signed off.

That name stays there, out of sight, while everything around it moves on. AI providers refresh their models often and retire the old ones with a notice period. OpenAI's own page has retirements scheduled on dates running from October 2026 to April 2027. Anyone building with AI is building on a part with a use-by date, even if they cannot see it.

The cost of finding out too late

The problem is not switching models, which is often straightforward. The problem is finding out on the day it stops working. At Elena's distributor, that could be a Monday in January with orders missing from the sheet, deliveries waiting to be packed and nobody sure where to start looking.

And even when the switch happens in time, a new model does not always behave like the old one. It might read an order differently, take longer or cost a different amount. If nobody tests it on real emails, a mistake can go unnoticed for weeks, quietly logging the wrong quantities.

What should you do about the GPT-4o retirement?

If anything in your business runs on AI, you can work through these steps calmly before January:

• List everything that uses AI behind the scenes: the chatbot on your website, email automations, reports that build themselves, links to your management software.
• Find out which model each one uses. Whoever built it will know, and it also shows up in your company's OpenAI account dashboard.
• Check that the account belongs to someone on the team and that OpenAI's notices reach an inbox someone actually reads.
• Gather twenty or thirty real cases, including the awkward ones: the order full of typos, the one mixing two customers, the one sent as an attachment.
• Test the replacement on those cases and compare its answers with the current model's before switching over for day-to-day work.
• Note down which model each tool uses and a date to review it, so the next retirement does not catch you out.

At the distributor, the test could be as simple as running a month of orders through the new model and comparing the resulting sheet with the one they already have. If it matches, they switch. If not, there is time to adjust the instructions before January.

How we help at digizone

When we take on web app and bespoke software development, the parts that run on their own are part of the product from day one: they are documented, they live in the client's own infrastructure and they can be reviewed and changed without rebuilding anything. And if you already have an automation someone else built, we can go through it with you, find out which model it uses and plan the switch calmly.

For WannaNow, we built an importer that trawls the city's public listings every day, standardises each event, categorises it and places it on the map with no human involvement. We built it on Supabase, with scheduled tasks inside the database itself. That way, the job that runs on its own every day lives inside the system itself.

Start with a list of your automations

You do not need to rebuild anything to have peace of mind. The first version of this job is a simple list: which tools use AI, which model each one runs on, who looks after it and when it gets reviewed. At Elena's distributor, it would be a single row for the order automation. With that list in place, January's retirement becomes just another date in the diary.

If you would like to see how it would work for you, our free workshop, in Spanish, walks through it step by step.