GPT-4o is switched off on 14 January 2027. Here's what that means for your business, and a 12-week migration plan

On 14 October 2026, OpenAI told developers that four of its most widely used models will be removed from the API on 14 January 2027: GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini. A lot of the AI built in the last two years runs on one of them, including customer-service bots, document tools, internal assistants and AI features inside SaaS products.

On that date, requests to those models stop working. OpenAI's own definition is short: at shutdown, "the model or endpoint will no longer be accessible."

We're moving our own product, ConGreeto, off these models, and we've done the testing this post describes. This is what we learned. It's written for the person who has to make sure nothing breaks in January. That might be a founder, an operations lead, or a CTO with a system someone else built.

What's being switched off, and when

From OpenAI's deprecations page, checked 22 October 2026:

Model

Snapshots named in the notice

Shutdown

OpenAI's recommended replacement

GPT-4o

gpt-4o-2024-08-06, gpt-4o-2024-11-20

14 Jan 2027

gpt-6-luna

GPT-4o mini

gpt-4o-mini-2024-07-18

14 Jan 2027

gpt-6-luna

GPT-4.1

gpt-4.1-2025-04-14

14 Jan 2027

gpt-6-luna

GPT-4.1 mini

gpt-4.1-mini-2025-04-14

14 Jan 2027

gpt-6-luna

Fine-tuned models go too. In OpenAI's words: "Fine-tuned versions of these models follow the same shutdown schedule as their underlying base models." A separate fine-tuning notice from May 2026 says that from 6 January 2027, existing customers can no longer create new fine-tuning jobs. If your system depends on a fine-tune of one of these models, you probably can't fix it by fine-tuning again. Plan to replace the fine-tune with better prompts, examples or retrieval, and check what the May notice means for your account specifically.

Three nearby dates catch people out:

• 20 January 2027: legacy realtime and audio models are switched off, including gpt-4o-realtime and gpt-4o-audio. This affects voice agents.
• 26 February 2027: older transcription models are switched off, including whisper-1 and gpt-4o-transcribe, along with the /v1/audio/translations endpoint. This affects anything that turns calls or meetings into text.
• 26 August 2026 (already passed): the Assistants API was shut down. If something of yours stopped working in late August, that's probably why.

Are you affected?

Not affected: your staff using the ChatGPT app or website. This notice is about the developer API, which is what software calls behind the scenes.

Affected: any software that calls one of those models through the API. It might be code your team wrote, a system an agency built for you, or a third-party tool that uses OpenAI under the hood.

How to find out, in roughly an hour:

1. Search your code and configuration for gpt-4o and gpt-4.1. Check environment variables, settings files, database tables and admin screens too, because model names often live outside the code. Look for the short alias (gpt-4o) as well as dated snapshots.
2. Check your OpenAI usage dashboard. It shows which models your account has called recently, which catches systems nobody remembers building.
3. Ask your vendors in writing. "Does your product call GPT-4o, GPT-4o mini, GPT-4.1 or GPT-4.1 mini, and what's your migration date?" Any AI tool you pay for might be affected.
4. If you use Azure OpenAI, check Microsoft's own model retirements page. Azure sets its own dates per model version and region. They don't follow OpenAI's.

Why it isn't a find-and-replace

The tempting plan is to change gpt-4o-mini to gpt-6-luna and redeploy. Three things make that risky.

1. The replacement is a reasoning model, and its default setting isn't the one you want

GPT-6 Luna can "think" before it answers, and you choose how much with a reasoning effort setting. OpenAI's model page lists six levels, from none to max, with medium as the default.

OpenAI's reasoning guide says reasoning tokens "are not visible via the API" but are billed as output tokens and count towards your output limit. It also says "lower effort favors speed and lower token usage."

So if you swap the model name and change nothing else, each reply runs at medium effort. It will be slower than the GPT-4o mini reply it replaces, and part of what you pay for is thinking nobody sees. For a chatbot or a classification step, that's usually the wrong trade.

2. The same instructions can produce different behaviour

On 15 October 2026 we tested the replacement models on our own product's real prompts: two live assistants, 22 test cases, each run at several effort settings. A small internal test, not a benchmark, but every result was something we'd have shipped to customers:

• At none effort, GPT-6 Luna broke an instruction the old model followed. The assistant had been told to reply only in English. Asked in another language, it replied in that language anyway.
• At low effort, it followed the rule. So did GPT-6.1 Sol, OpenAI's more expensive option.
• GPT-6.1 Sol at medium gave the same answers as at low, about three times slower. At the 90th percentile, a reply took around 12 seconds. A website visitor won't wait that long.

The lowest setting was cheapest and fastest, and it was wrong. The most expensive setting was right, and too slow. The setting we chose wasn't obvious until we tested. You'll only find yours by running your own cases.

3. Testing finds problems that aren't the model's fault

During the same test, both old and new models quoted prices that were out of date. The cause wasn't the model. An old uploaded document was still feeding the assistant stale prices after the website had been updated. A migration is the first time many teams re-read their AI's answers carefully, so expect to find content and data problems as well as model problems. Budget time to fix them.

What happens to your bill

On paper, GPT-6 Luna is far cheaper. Standard prices from OpenAI's pricing page, per million tokens, checked 22 October 2026 (short-context rates):

Model

Input

Cached input

Output

GPT-4o

$2.50

$1.25

$10.00

GPT-4.1

$2.00

$0.50

$8.00

GPT-4.1 mini

$0.40

$0.10

$1.60

GPT-4o mini

$0.15

$0.075

$0.60

GPT-6 Luna

$0.10

$0.01

$0.50

GPT-6.1 Sol

$2.00

$0.10

$10.00

Moving from GPT-4o or GPT-4.1, the per-token saving is large: input is 20 to 25 times cheaper. Most systems will cost much less to run.

Moving from GPT-4o mini, the saving is small, and reasoning can wipe it out. An illustrative reply, with numbers chosen for easy arithmetic rather than measured from any system: 4,000 input tokens, 300 visible output tokens, no caching.

Input cost

Output cost

Per reply

Per 100,000 replies

GPT-4o mini

$0.00060

$0.00018

$0.00078

$78

GPT-6 Luna, no reasoning

$0.00040

$0.00015

$0.00055

$55

GPT-6 Luna, 500 hidden reasoning tokens

$0.00040

$0.00040

$0.00080

$80

The same model can be cheaper or dearer than the one it replaces, depending on one setting. Measure real token counts during testing, including reasoning tokens (the API reports them separately), before you promise anyone a lower bill.

Caching matters even more on the new model. Luna's cached input is a tenth of its normal input price. If your prompts start with a long, unchanging block (instructions, product details, policies), keeping that block identical and at the start of the prompt can cut input cost sharply. We cover how in what an AI agent actually costs to run.

A 12-week migration plan

From today, 22 October 2026, there are about twelve weeks until shutdown. Two of those fall over the Christmas and New Year period, when few teams want to ship changes. Plan to finish by mid-December.

Week

Step

Output

1

Inventory. Find every call to the four models, including vendors and Azure deployments.

A list: system, model, owner, business impact if it stops

1–2

Build an evaluation set. 20–50 real inputs per system with the answer you expect, including the awkward ones: rule-breaking requests, other languages, missing information.

A test you can re-run in minutes

2

Record the baseline. Run the set on the current model and save the answers, cost per reply and response time.

Your "before" numbers

3–4

Test the replacement at two or three effort levels. Start with low, and compare none and medium. Try the larger model only where quality falls short.

Pass rate, cost and response time for each setting

4–5

Fix what broke. Usually prompt wording, output format checks, timeouts, and the stale content testing turns up.

Evaluation set passing at the chosen setting

6

Make the model a setting, not code. Model name and effort level should be configurable per feature, so the next migration is a config change and a test run.

One place to change models

6–7

Roll out gradually. Send a small share of real traffic to the new model, compare, then increase.

Live comparison on real users

8

Switch fully, and keep the rollback. Rolling back to the old model only works until 14 January, so finish early enough for it to be useful.

Production on the new model

9–12

Watch and buffer. Monitor cost, response time and complaints. This is your slack for anything that slipped.

Done before the holiday freeze

If you only have time for one step, do the evaluation set. Without it, you're deciding by trying ten questions by hand and hoping.

If someone else built your system

Many businesses running GPT-4o didn't build the system themselves. If that's you:

• Ask the builder now, in writing, whether they're handling the migration, by when, and at what cost.
• Ask for the evaluation results, not just "it's done". You want to see the test cases, the pass rate and the new cost per reply.
• Check you have access to the code, the prompts and the OpenAI account. If the original builder has gone, you'll need all three for anyone else to do the work.
• Use the migration to fix the contract. This won't be the last model retirement. Agree now who pays for the next one. We list that and eleven other questions in what to ask an AI development company before you sign.

This will happen again

Every major model provider retires models. In 2026 alone, OpenAI has already shut down chatgpt-4o-latest, the Assistants API and its realtime previews. Treat this migration as the chance to make the next one cheap:

• Model and settings in configuration, per feature.
• An evaluation set that runs before every change.
• Cost and latency recorded for every call, so you can see the effect of a change in days, not on next month's invoice.
• Code that isn't tied to one provider's special features unless they're worth it, so you can switch providers when price, quality or terms change.

With those in place, a deprecation notice becomes a morning's work and a test run instead of a project.

Frequently asked questions

When is GPT-4o being shut down?

OpenAI removes GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini from its API on 14 January 2027, according to its deprecations page (notice dated 14 October 2026). Fine-tuned versions of those models are switched off on the same date.

Does the GPT-4o shutdown affect ChatGPT?

This notice covers the developer API, the way software calls OpenAI's models. It doesn't describe changes to the ChatGPT app. If your business only uses ChatGPT in the browser, this notice doesn't require you to do anything.

What should I replace GPT-4o with?

OpenAI recommends gpt-6-luna for all four retiring models. It's a reasoning model with a configurable effort level, and its default (medium) is often slower and more expensive than a chatbot needs. Test it on your own cases at low and nearby settings before choosing.

Will migrating from GPT-4o reduce my costs?

Coming from GPT-4o or GPT-4.1, almost certainly: Luna's listed input price is 20 to 25 times lower. Coming from GPT-4o mini, the saving is small, and hidden reasoning tokens, billed as output, can cancel it. Measure real token usage during testing.

What happens if I don't migrate in time?

Calls to the removed models stop working on the shutdown date. Whatever depends on them, such as a chatbot, a document workflow or an AI feature in your product, will fail or fall back to whatever error handling you have.

How long does a GPT-4o migration take?

For a single, well-understood feature with a test set, days. For a system with several AI steps, no test set and a builder who has moved on, plan for several weeks. Most of the time goes into testing, not into changing the model name.

Sources

• OpenAI, Deprecations: GPT-4 series notice dated 14 October 2026 (shutdown 14 January 2027); legacy audio and realtime notice of 20 July 2026 (shutdown 20 January 2027); transcription notice of 26 August 2026 (shutdown 26 February 2027); fine-tuning notice of 7 May 2026; Assistants API shutdown 26 August 2026. Checked 22 October 2026.
• OpenAI, GPT-6 Luna model page: reasoning effort levels and default, checked 22 October 2026.
• OpenAI, Reasoning guide: reasoning tokens billed as output tokens, checked 22 October 2026.
• OpenAI, API pricing: standard short-context rates, checked 22 October 2026.
• Microsoft, Azure OpenAI model retirements.
• Model-replacement test: Brillnex Systems, two ConGreeto assistants, 22 test cases, 15 October 2026. A small internal test, not a benchmark.

From us

• 12 questions to ask an AI development company before you sign
• What an AI agent actually costs to run in production
• What a RAG chatbot actually costs: build, buy or have it built
• AI agent development: if you have a system on GPT-4o and no clear owner for the migration, tell us what it does. We'll reply in writing with what needs to change and roughly how long it will take, before any call.