GPT-4o leaves the API in January: how to switch the model behind your automations without surprises

What OpenAI announced

According to OpenAI's deprecations page, these are the models and snapshots shut down on January 14, 2027. For all of them, OpenAI recommends the same replacement: GPT-6 Luna, released in September at US$0.10 per million input tokens and US$0.50 per million output tokens.

• GPT-4o (snapshots 2024-08-06 and 2024-11-20).
• GPT-4o mini (snapshot 2024-07-18).
• GPT-4.1 (snapshot 2025-04-14).
• GPT-4.1 mini (snapshot 2025-04-14).
• Fine-tuned models built on them follow the same schedule and will also be shut down.

The same week, more options on the market

The notice came in a busy week. On October 12, Meta released Muse Wonder, its first frontier model, at US$4 per million input tokens and US$20 per million output tokens, with performance comparable to OpenAI's and Anthropic's most advanced models in Meta's own testing. On October 13, Google made Gemini 4 Argon available to developers and enterprises, at a promotional price of US$2 and US$10 until November 30, rising to US$4 and US$20 afterward.

Also that week, Replit CEO Amjad Masad said on the TechCrunch Disrupt stage that the platform now has OpenAI as its preferred partner instead of Anthropic, citing cost, partnership and performance — and noted that users can still choose other models.

The message for anyone automating is clear: the model you use today will change, whether because the vendor shuts it down, a better option appears or the price changes. The question is whether the switch will be planned or rushed.

Why switching models is more than changing a name in the code

Technically, the switch can be a single line of configuration. In practice, each model interprets the same instructions a little differently, and what worked well on the old model may change on the new one:

• Response format: fields in a different order, longer or shorter text, JSON that no longer follows exactly what was expected.
• Tone of voice: customer service tuned to sound a certain way may become more formal, more verbose or more curt.
• Rule following: instructions such as "never quote a price" or "hand over to a human when the customer asks for a discount" need to be tested again.
• Tool use: the way the model decides to check inventory, calendars or the CRM may change.
• Fine-tuning: a model trained on your own data on top of GPT-4o mini does not migrate by itself; it has to be retrained or replaced with instructions and document retrieval.

A plan for the next three months

There is time to do this calmly, as long as the work starts now:

• Inventory: list every place these models are called — in-house systems, flows in tools such as n8n and Make, vendor chatbots and forgotten scripts.
• Test cases: gather real conversations and documents, including the hard cases, with the expected result for each one.
• Comparison: run GPT-6 Luna and at least one alternative against those cases, measuring quality, response time and cost.
• Instruction tuning: fix whatever changed in behavior before going to production.
• Gradual rollout: start with part of the volume, track the metrics and only then migrate everything.
• Vendors: if the automation is a third party's, ask which model it uses and what the migration plan is.

Design for the next switch

This will not be the last deprecation. It is worth using the migration to make the model configurable, separate from the business logic, and to keep the test case set up to date. That way, the next switch — whether forced or by choice, such as after a price drop — becomes a routine of a few days rather than a project.

Frequently asked questions

What happens if I don't migrate by January 14, 2027?

Calls to the shut-down models stop working. In practice, the automation stops: customer service doesn't reply, documents aren't processed. That is why it is best to finish the switch weeks ahead of time, to leave room for adjustments.

Do I have to use the replacement OpenAI recommends?

No. GPT-6 Luna is OpenAI's recommendation, but the switch is a good opportunity to compare alternatives from other vendors. The criterion should be performance on your real cases, together with cost and response time.

Are other OpenAI models being retired too?

Yes. GPT-4.1 nano already had a shutdown date of October 23, 2026, and audio and realtime models have their own date, January 20, 2027. It is worth checking OpenAI's deprecations page for every model you use.