GPT-4o Retires Jan 14: What Breaks in Your Business

Somewhere in your business, a workflow calls an AI model by name. Odds are you didn't type that name, and you can't recite it. On October 14, 2026, OpenAI notified developers that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini leave the API

Somewhere in your business, a workflow calls an AI model by name. Odds are you didn't type that name, and you can't recite it.

On October 14, 2026, OpenAI notified developers that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini leave the API on January 14, 2027. This GPT-4o retirement has a long tail, because those models sit under chatbots, n8n flows, quote generators and voice agents built over the last two years.

The headline date is not the first date

Most coverage stops at January 14. OpenAI's deprecations page lists a whole calendar, and one entry lands this week.

October 23, 2026 (this Friday): gpt-3.5-turbo, gpt-4, gpt-4-turbo, gpt-4.1-nano, the gpt-4o-2024-05-13 snapshot, o1, o3-mini, o4-mini and gpt-image-1. Fine-tuned versions of several of these go too.

December 11, 2026: the original GPT-5 snapshots, including the mini and nano versions.

January 6, 2027: the tts-1 and tts-1-hd text-to-speech models, plus the gpt-4o-mini-tts snapshots.

January 14, 2027: GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini.

January 20, 2027: the older realtime and audio families, including gpt-realtime and gpt-4o-realtime.

February 26, 2027: whisper-1 and the GPT-4o transcription models.

The notice period is shorter than the promise

OpenAI's page says generally available models get "at least 6 months" of notice unless safety or compliance forces a faster timeline. Count from October 14 to January 14 and you get three.

The page doesn't explain the gap. Maybe these models were flagged as legacy earlier, maybe not. Either way, plan against the dates in the table, not against the policy paragraph above it.

Software relying on OpenAI models may need occasional updates to keep working.

— OpenAI, API deprecations page

That sentence is the whole business case for treating AI plumbing as maintained infrastructure, not a one-off project.

Old model names hide in five places

You are looking for any tool where a human once typed a model name. Here is where they end up.

Automation flows. An n8n, Make or Zapier scenario with an OpenAI step that has a model dropdown or text field set to gpt-4o-mini because that was the cheap default in 2025. We compared where each platform fits in our n8n vs Claude agents breakdown. Any of them can hold a retiring name.

Website and support chatbots. A freelancer wired the chat widget to the API with a model string in the code. Nobody has opened that file since launch.

Voice agents. Our Retell vs Vapi vs Bland cost comparison describes a Vapi setup tuned around GPT-4o-mini as the language model. The voice layer is on the list too: text-to-speech, realtime and transcription models all have shutdown dates between January 6 and February 26. If you want the baseline for what a phone agent does for a small business, start with our AI phone receptionist guide.

Fine-tuned models. OpenAI says fine-tuned versions follow their base model's schedule. It also says active existing customers can't create new fine-tuning jobs after January 6, 2027. A fine-tuned GPT-4o mini is a rebuild on a new base, with a fresh evaluation, not a toggle.

Scripts nobody owns. The spreadsheet add-on, the invoice-summary script, the intern's Python file that still runs every Monday. These fail quietly, and quiet failure is the expensive kind.

The replacement is cheaper, which is the trap

OpenAI names GPT-6 Luna as the recommended replacement for all four January 14 models. Its pricing page lists standard short-context rates per million tokens, and the gaps are large.

GPT-4o mini: $0.15 input, $0.60 output. GPT-4.1 mini: $0.40 and $1.60. GPT-4.1: $2.00 and $8.00. GPT-4o: $2.50 and $10.00. GPT-6 Luna: $0.10 and $0.50.

So the bill almost certainly falls. That is the good news, and it is also how people talk themselves out of testing.

One model replacing four, across a 25x price spread, means OpenAI is not claiming these are drop-in equivalents. If a task ran on GPT-4o for its judgment, such as classifying an angry email or deciding whether a lead is qualified, Luna may do it fine. It may also be softer on the edge cases. The same page lists GPT-6 Sol at $2.00 and $10.00, close to GPT-4o's old bracket, which is where the heavier tasks may need to go. We haven't benchmarked either against your data, and neither has anyone else.

The token bill is pocket change either way

Here is a worked example, with invented usage numbers so you can swap in your own. Say a website chatbot handles 600 conversations a month, and each one burns about 3,000 input tokens and 500 output tokens. That is 1.8 million input tokens and 0.3 million output tokens monthly.

At the listed standard rates, GPT-4o mini costs about $0.45 a month for that traffic. GPT-6 Luna costs about $0.33. Even the full GPT-4o comes to roughly $7.50.

Read that twice. Moving off GPT-4o saves a small business the price of a sandwich. The reason to migrate carefully has nothing to do with the invoice.

The risk sits on the other side of the ledger: one wrong answer to a customer who was about to book. Test time is the real cost, so spend it where the conversations are worth the most.

Prompts drift before invoices do

A new model reads the same prompt differently. The failure isn't an error message. It's a chatbot that suddenly apologizes twice per answer, a lead scorer that now rates everything 7 out of 10, or a receptionist that books the wrong appointment type.

Here is the rule we would hold any build to, and it costs nothing to adopt. The model name lives in one config field, never scattered through prompts and nodes. And the workflow has a saved set of real inputs, say twenty emails, calls or form fills, that you can replay against the old and new model and compare side by side.

Without that test set, "we migrated" means "it didn't crash."

A ninety-minute audit for this week

You don't need an engineer for the first pass. You need a list.

1. Write down every tool that touches AI. Chatbot, automations, receptionist, content tools, scripts. Include the ones a contractor set up.
2. Pull your OpenAI usage by model. If your business has an API account, the usage dashboard shows which models are being billed. Anything with a retiring name is a finding.
3. Ask each vendor three questions in writing. Which model do you call for us, which date does it change, and who tests the new one? A rented tool should handle the swap on its side. You want that promise on record.
4. Use the shutdown field. OpenAI documents a shutdown_date on its List models endpoint, which a script can check on a schedule so the next retirement doesn't arrive by email.
5. Check who gets OpenAI's email. OpenAI says it notifies customers actively using a model by email. If that address belongs to someone who left in March, the notice went nowhere.
6. Book the cutover before the holidays. A January 14 deadline means testing in the first week of January, when half your team is out. December is the sane month.

Rented tools migrate on the vendor's clock, built tools on yours

If you pay a platform for a receptionist or chatbot, the model choice is mostly theirs. Your job is to ask the questions above and hold them to the answers.

If someone built the thing on the raw API, nobody is coming. The work isn't editing a string. It's finding every call, replaying real cases, tuning prompts to the new model, and watching the first week of live traffic.

That is the kind of job we do under our automation work: inventory the workflows, move them, and leave you with a test set so the next retirement notice is an afternoon, not a scramble. We are a vendor, so weigh that. The audit above is yours to run for free, and plenty of small teams will finish it themselves.

The same pattern showed up last month with custom GPTs being retired on December 11. Different product, same lesson: anything you build on a vendor's menu has an expiry date.