The GPT-4o Shutdown Is January 14: Your CMS AI Plugin May Run on GPT-4o Mini

The GPT-4o shutdown is set for January 14, 2027. On October 14, OpenAI posted a notice that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini will be removed from the API on that date. Fine-tuned versions of those models stop on the same day. The named replacement for all of them is GPT-6 Luna.

These four models have been in the API since 2024 and early 2025. If a WordPress or Drupal AI feature, a chatbot or a Zapier step was set up in 2024 or 2025 and never touched, check it. This post covers what is on the list, where the model IDs hide on each platform, what changes on GPT-6 Luna, what happens to fine-tunes, and a calendar that runs through the holidays. Facts are as of October 15, 2026.

The short version: you have about three months, and the last few weeks fall over Christmas and New Year. GPT-6 Luna costs less per token than all four models. It is also a reasoning model that thinks by default, which can raise cost, slow answers and break function calling in older plugin code. Fine-tunes have no OpenAI base model left to retrain on. Find your model IDs this month.

What the GPT-4o shutdown covers

The entry on OpenAI’s deprecations page is headed "GPT-4 series model deprecations." It lists five dated snapshots.

The entry says: "Fine-tuned versions of these models follow the same shutdown schedule as their underlying base models."

The bare names are not in the table. Most plugins call gpt-4o-mini or gpt-4.1-mini, not a dated snapshot. OpenAI’s model pages now mark all four models as Deprecated, and they do not say which snapshot each bare name points to. Plan as if the bare names stop on January 14 too.

The notice is shorter than usual. OpenAI’s policy on the same page promises at least six months for generally available models. October 14 to January 14 is three. The entry gives no reason.

This wave comes after the October 23 shutoff, which took the older gpt-4o-2024-05-13 snapshot and GPT-4.1 nano, and before the April 1 shutoff of GPT-5.1 and GPT-5.4 nano.

ChatGPT lost GPT-5.5 the same day

A separate change landed on October 14. OpenAI’s model release notes say: "We’re retiring GPT-5.5 from ChatGPT, ChatGPT Work, and Codex." The note adds: "This change does not affect API access."

Staff using ChatGPT lost GPT-5.5 but keep working. API integrations on GPT-4o keep working until January 14, then fail.

If a client asks whether the ChatGPT change affects their website, the answer is no. The GPT-4o shutdown is the one that can.

Where the GPT-4o shutdown hides in code and databases

Start with OpenAI’s own usage data grouped by model and API key. Our April shutoff post shows how. Then search for these strings:

gpt-4o
gpt-4o-mini
gpt-4.1
gpt-4.1-mini

In a code checkout, one search covers all four:

grep -rnE "gpt-4o|gpt-4\.1" --exclude-dir=node_modules .

Expect noise. gpt-4o also matches gpt-4o-mini-tts, gpt-4o-transcribe and the realtime models. Those are on the same deprecations page with their own dates in January and February, so note them while you are there.

Search the database, not only the code. Many CMS plugins save their model choice as a setting. On WordPress, WP-CLI can search every table:

wp db search "gpt-4o"
wp db search "gpt-4.1"

On Drupal, export the active configuration to a scratch folder and search it. That catches settings that were never exported to the repository:

drush config:export --destination=/tmp/config-check
grep -rnE "gpt-4o|gpt-4\.1" /tmp/config-check

Check hosting dashboards. A model set in an environment variable on a serverless function will not appear in either search.

WordPress, Drupal, automations and chatbots

Each platform puts the model choice in a different place.

AI Engine (WordPress). The AI Engine plugin lets you pick a model per chatbot and per environment. Its changelog shows version 3.8.4 added GPT-6 Sol, GPT-6 Luna and GPT-6.1 Sol. Update the plugin first, then check the default model and every chatbot.

The WordPress AI plugin. The official AI plugin on WordPress.org connects providers through Settings, Connectors. Its listing does not describe a model picker. Version 1.3.0 "Updated preferred models to more recent ones for the three default providers." Here the plugin chooses, so keeping it updated is the fix.

Drupal AI module. The AI module documentation says you set a default model for each operation type at /admin/config/ai/settings. It also says most contrib modules can override that default. Check the defaults, then each AI module. If the site is still on Drupal 10, support ends December 9, 2026, before this shutdown. See our Drupal 10 end-of-life guide and our plain-language Webform vulnerability brief.

Zapier and Make. List every Zap and scenario with an OpenAI or ChatGPT step. Open each one and note the model it uses.

Hosted chatbots. If a vendor runs the chatbot, you cannot see the model. Ask the vendor in writing which model it uses and when it will move.

What changes when a plugin moves to GPT-6 Luna

On price, the GPT-4o shutdown works in your favor: GPT-6 Luna is cheaper than all four retiring models. GPT-6 Luna prices come from the OpenAI pricing page. The older models no longer appear there, so their prices come from their model pages. All prices are per million tokens.

From GPT-4o mini, input drops by a third and output by about 17%. From GPT-4o or GPT-4.1, both drop by about 95%. Luna adds a cache write charge of $0.125, and Batch runs at half price.

The GPT-6 Luna model page lists a 1,050,000-token context window and 128,000 output tokens. GPT-4o mini had 128,000 and 16,384.

Rate limits depend on your usage tier. On the lowest tier, Build, Luna allows 5,000 requests and 2 million tokens per minute. Our post on OpenAI usage tiers covers what to check per client account.

Reasoning is the change that breaks things

GPT-4o mini answered directly. GPT-6 Luna is a reasoning model, and its default effort is medium. Change only the model ID and every request starts thinking first. Three things follow.

Cost. Reasoning tokens bill as output, as we explained in input vs output tokens. A cheaper rate can still produce a bigger bill.

Speed. OpenAI’s reasoning guide describes none as the setting for latency-critical tasks. A chat widget that felt instant on GPT-4o mini may lag at medium.

Short output limits. The same guide recommends reserving at least 25,000 tokens for reasoning and output. With less room, you can pay for input and reasoning and get no visible answer. Plugins that cap output at a few hundred tokens for alt text or summaries are at risk.

Function calling. The Luna model page says: "Chat Completions supports function calling only with reasoning_effort set to none." Plugin code built on Chat Completions with tools must set none or move to the Responses API.

For most jobs that ran well on GPT-4o mini, start testing Luna at none. Then try low where answers fall short. OpenAI names Luna for GPT-4o and GPT-4.1 too. For work that used those larger models for judgment, also test GPT-6.1 Sol. Our model selection guide compares the tiers, and our GPT-6 Sol and Luna launch post covers Luna’s benchmark results.

Fine-tuned models have nowhere to go on OpenAI

Fine-tunes built on these four models stop at the GPT-4o shutdown on January 14. The question is what to retrain on.

OpenAI’s fine-tuning guide lists the base models you can still tune. Supervised and preference tuning use GPT-4.1, GPT-4.1 mini and GPT-4.1 nano. Reinforcement tuning uses o4-mini-2025-04-16. Vision tuning uses gpt-4o-2024-08-06.

Every one of those is shut down on October 23 or January 14. The GPT-6 Luna model page lists fine-tuning as not supported. Separately, the deprecations page says active existing customers can no longer create fine-tuning jobs after January 6, 2027.

Your training data is still useful. Put 20 to 50 of the best examples in the system prompt and compare results against the fine-tune before the GPT-4o shutdown. Prompt caching keeps a long example block cheap.

A GPT-4o shutdown calendar that runs through the holidays

Thanksgiving and year-end closures take a share of the three months. Here is one way to lay out the GPT-4o shutdown work.

1. By October 30: finish the October 23 work. Pull usage by model and API key, run the searches above, and list every site, plugin, Zap and chatbot on the four models.
2. By November 13: save 50 real prompts and current answers per workload. Ask plugin and chatbot vendors when they support GPT-6 Luna.
3. By November 25: test Luna at `none` and `low`. Compare answers, length, speed and cost. Decide how to replace each fine-tune.
4. November 30: Claude Sonnet 4.5 retires, along with OpenAI’s prompts API, Evals and Agent Builder.
5. By December 11: move background jobs such as tagging and alt text. The GPT-5 snapshots shut down that day.
6. By December 18: move chatbots and anything members see, while staff are still in the office.
7. December 21 to January 1: no changes. Watch error logs.
8. January 6: fine-tuning job creation ends, and the `tts-1` models shut down.
9. By January 13: confirm zero usage on the old IDs and remove them from settings. Alert on failed requests, not on one error code.

How Anthropic and Google handle the same problem

The GPT-4o shutdown is not the only retirement this fall. Each vendor handles it differently.

Anthropic retires claude-sonnet-4-5-20250929 on November 30, 2026, with claude-sonnet-5-5 as the replacement. Its deprecations page promises at least 60 days’ notice for publicly released models. Our Claude Sonnet 5.5 post covers the breaking changes.

Google took another route on October 8. Its Gemini deprecations page says requests to gemini-3.7-flash now go to gemini-3.8-flash, and gemini-3.5-flash goes to gemini-3.6-flash. The page says the endpoint "is not shutting down." Nothing fails, but a different model answers. Our post on Gemini 4 Argon going generally available covers what that does to cost.

For OpenAI and Anthropic, the check is a search for the ID before the date. For Google, it is a look at billing and output since October 8.

Frequently Asked Questions