OpenAI retires GPT-4o and GPT-4.1 in the API: moving to GPT-6 Luna

Computer processor held by tweezers against a grey background

Two dates that affect your integrations

If your application, chatbot, n8n automation or document-processing script calls the OpenAI API, you will almost certainly have to change models within the next three months. On October 14 OpenAI notified customers that on January 14, 2027 it will remove popular GPT-4 series models from the API. On top of that, tomorrow, October 23, it shuts down a set of older models announced back in April.

October 23, 2026: the following stop working, among others:

gpt-3.5-turbo, gpt-4, gpt-4-turbo, gpt-4-1106-preview;
gpt-4o-2024-05-13 and gpt-4.1-nano;
o1, o1-pro, o3-mini, o4-mini;
gpt-image-1 (image generation);
fine-tuned models built on some of them, e.g. ft-gpt-3.5-turbo and ft-gpt-4.

Replacements for this group are GPT-5.6 family models (gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna) and gpt-image-2.5 in two variants.

January 14, 2027: gpt-4o (snapshots 2024-08-06 and 2024-11-20), gpt-4o-mini, gpt-4.1 and gpt-4.1-mini go away. OpenAI names a single replacement for all of them: gpt-6-luna. Fine-tuned models based on them are shut down the same day and have to be retrained or dropped.

Shorter notice than OpenAI promises

Under OpenAI's own policy, generally available models are retired with at least six months' notice unless safety or compliance concerns require a faster timeline. For GPT-4o and GPT-4.1 the notice is 92 days, and OpenAI has not given a reason for the shorter period. MIXED and The New Stack pointed this out. For businesses it means the migration has to fit around the holidays and year-end close, so it is better to start now.

The list covers the API. ChatGPT as used by your staff is a separate matter and is not part of this notice.

GPT-6 Luna: cheaper, but a different model

According to OpenAI's pricing, gpt-6-luna costs USD 0.10 per million input tokens, USD 0.01 for cached input and USD 0.50 per million output tokens. For comparison, GPT-4o is 2.50 / 1.25 / 10 USD. The context window grows from 128k to 1.05 million tokens, and the maximum response is 128k tokens. A few caveats that affect the bill:

Luna is a reasoning model. The default reasoning.effort is medium, and reasoning tokens are billed as output. The same prompt may therefore use more tokens than on GPT-4o.
Requests with more than 272k input tokens are billed in full at 2× the input rate and 1.5× the output rate.
Batch and Flex cost 50% of the standard rate; Fast mode costs twice as much.
Luna does not support fine-tuning. If your process relies on a fine-tuned GPT-4o, you need to move it to a prompt with examples or to another model.
In Chat Completions, function calling works only with reasoning_effort set to none. Full tool support is in the Responses API.
OpenAI describes Luna as a model "for focused, high-volume tasks". For complex analysis, also consider the more expensive GPT-6 models and compare results.

Personal data, GDPR and the AI Act

Luna supports EU data residency in Standard, Fast, Flex and Batch modes. Regional processing costs 10% more. If your OpenAI project is configured for EU data residency, check that the new model and the chosen mode are available there before switching production traffic. Your data processing agreement with OpenAI does not change, but a model change is a change of a processing component: record it in your documentation and records of processing. If a system using an OpenAI model supports decisions about people (recruitment, customer assessment), the AI Act requires you as the deployer to oversee and monitor its operation, so results on the new model must be checked before rollout.

How to check whether this affects you

Search code and configuration for model names, e.g. grep -rnE "gpt-4o|gpt-4\.1|gpt-3\.5|gpt-4-turbo|o1|o3-mini|o4-mini|gpt-image-1" .
Remember places outside the repository: environment variables, n8n, Make, Zapier, WordPress plugins, CRMs and helpdesks with a built-in OpenAI integration.
In the OpenAI dashboard (Usage), check which models actually consume tokens in each project.
The GET /v1/models endpoint returns a shutdown_date field. You can check it automatically and alert on upcoming shutdowns.
If you use Azure OpenAI, Microsoft sets its own retirement schedule. Check it for your deployments.

How to migrate without surprises

Collect 30–100 real (anonymised) production requests and the expected answers.
Run them through gpt-6-luna at several reasoning levels, from none to medium. Compare quality, response time and cost.
Check format compatibility: JSON with a schema (structured outputs), function calls, response length.
Switch part of the traffic, watch errors and costs, then move the rest. Leave yourself a buffer before January 14.

Other shutdowns are queued as well: on November 30 Agent Builder, the Evals platform and the reusable prompts API go away, on December 11 older GPT-5 and o3 snapshots, and on January 6 the tts-1 models. A model is a dependency with an expiry date: keep its name in configuration, maintain an inventory of where it is used, and have a test set that lets you swap it in a day.