OpenAI is retiring GPT-4o on 14 Jan 2027: what breaks in your n8n/Make/Zapier automations and how to fix it

If you built an automation in the last two years that "uses ChatGPT", there is a good chance it calls gpt-4o, gpt-4o-mini or gpt-4.1-mini. OpenAI shuts those down on 14 January 2027. After that date the OpenAI step returns an error, and whatever sat downstream (the reply to a lead, the row in a sheet, the summary in Slack) stops arriving.

This is a checklist for people who run automations in n8n, Make or Zapier and don't want to read API docs. All dates below come from OpenAI's deprecations page (checked 14 Nov 2026). Check that page again before you act, because OpenAI adds to it.

The dates that matter

Two details people miss:

• Fine-tunes die with their base model. OpenAI's page says fine-tuned versions follow the same shutdown schedule as the base model. A ft:gpt-4o-mini-... model stops on 14 Jan 2027 too.
• You may not be able to retrain. On 6 Jan 2027, even active existing customers lose the ability to create new fine-tuning jobs. If your workflow depends on a fine-tune, plan the replacement now, not in the second week of January.

Already-dead models are worth a look too: gpt-3.5-turbo, gpt-4, gpt-4-turbo, o1, o3-mini and o4-mini shut down on 23 Oct 2026. If an old workflow has been silently failing since then, this is a good moment to find it.

Step 1: find every place a model name is set

n8n

Model names hide in more places than the OpenAI node:

1. OpenAI Chat Model sub-nodes attached to AI Agent, Basic LLM Chain or Q&A chains: check the Model field.
2. OpenAI node (text, image, audio): check the Model field.
3. HTTP Request nodes that call api.openai.com: the model is inside the JSON body, e.g. "model": "gpt-4o".
4. Code nodes that call the API directly.
5. Defaults you never chose. n8n doesn't save a parameter you left at its default, so the model name may not appear anywhere in the exported JSON. In current n8n, the OpenAI Chat Model default is gpt-5-mini, which retires on 11 Dec 2026. Older n8n versions defaulted to gpt-4.1-mini or gpt-4o-mini. Open the node and pick a model explicitly.
6. Transcription. The OpenAI node's audio transcribe operation is fixed to whisper-1 in n8n's source code, so you can't change it in the node. Before 26 Feb 2027, swap it for an HTTP Request node that calls the transcription endpoint with a supported model.

The fastest way to check everything is to scan all workflows at once. I built a free n8n workflow that does exactly that. It reads every workflow through the n8n API, finds retiring models and endpoints (including the default-model case above) and returns a table with the shutdown date, days left and the replacement. I've submitted it to the n8n template library. Until it's approved, you can export your workflows and use the scanner below.

Make

1. Open each scenario that has an OpenAI module and check the Model field in every OpenAI module.
2. Check HTTP modules that call api.openai.com and look for "model" in the request body.
3. To check a scenario in one pass, export its blueprint (the scenario's ⋯ menu, then Export Blueprint) and search the JSON file for gpt-, o3, whisper and tts.

Zapier

1. Filter your Zaps by the ChatGPT (OpenAI) app and open each step's Model field.
2. Check Webhooks by Zapier and API Request steps that call api.openai.com.
3. Steps that run on Zapier's own built-in AI (not your OpenAI API key) depend on Zapier's choices, not yours. They're worth a glance, but this checklist is about steps that use your key.

Custom code and small apps

Search your repo for gpt-4o, gpt-4.1, gpt-5, o3, whisper-1, tts-1, /v1/prompts and pmpt_. Don't forget environment variables and config files, where model names often live. You can also call GET /v1/models: OpenAI says each model can carry a shutdown_date field.

Step 2: test before you switch

The replacement is not a drop-in copy of the old model. Responses can differ in length, tone, how well they follow a JSON format, and speed. That matters most in automations, where the next step often parses the model's output.

A simple test that takes less than an hour:

1. Pick 5 to 10 real inputs your workflow handled recently. Include an awkward one.
2. Duplicate the workflow, change only the model, and run the same inputs through both versions.
3. Compare the outputs side by side. Check that every downstream step still works: JSON parsing, field mapping, message length limits, and any IF/Router conditions that read the AI's answer.
4. Keep the parameters the same and check that the new model accepts them. OpenAI's GPT-6 migration guide lists the differences.
5. Switch production only when the outputs match, and keep the old version for a day so you can compare live runs.

OpenAI's own advice is the same: run representative requests against both models and compare outputs, latency and cost before switching production traffic.

Step 3: put the date in your calendar

Set reminders a month and a week before each shutdown that affects you. Most of these are "it worked yesterday, it fails today" changes, with no gradual warning inside the automation itself.

Free tools

• OpenAI Deprecation Scanner (free): paste an n8n or Make export, a code file or a public GitHub repo. It lists every retiring model and endpoint with its location, shutdown date, days left and replacement.
• n8n template "Check your n8n workflows for retiring OpenAI models": scans your whole instance from inside n8n (currently in review for the n8n template library).

Disclosure: both are built by Fifty50 Labs, and so is the paid option below.

If you'd rather not do the migration yourself, our Migration Pack (from $99) switches your workflows to the replacement model and shows you before/after outputs on your own inputs.