OpenAI has put a date on the models a lot of small business automations were built on. On Jan. 14, 2027, five dated snapshots of GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini come out of the API, and OpenAI names gpt-6-luna as the replacement for all four (OpenAI deprecations).
That's 63 days from today. Nov. 12 to Nov. 30 is 18 days, December is 31, and January runs 14 more, so 18 + 31 + 14 = 63. Subtract the holidays and the honest number is smaller.
If a freelancer, a Zap or a Make scenario built your quote bot, inbox triage or support replies in the last two years, there's a decent chance one of these four names is sitting in a settings field nobody has opened since launch.
What OpenAI actually listed
The entry is dated Oct. 14, 2026. OpenAI says it "notified developers using certain GPT-4 series models of their deprecation and removal from the API on January 14, 2027" (OpenAI deprecations). The table names four families and five dated snapshots: GPT-4o (gpt-4o-2024-08-06, gpt-4o-2024-11-20), GPT-4o mini (gpt-4o-mini-2024-07-18), GPT-4.1 (gpt-4.1-2025-04-14) and GPT-4.1 mini (gpt-4.1-mini-2025-04-14). Then comes the line for anyone who paid to train a custom model:
Fine-tuned versions of these models follow the same shutdown schedule as their underlying base models.
There's no like-for-like rebuild, either. The gpt-6-luna model page lists fine-tuning as "Not supported" (OpenAI), and OpenAI's pricing page says "Fine-tuning is closed to new users" (OpenAI pricing).
Two details the headline skips. The entry lists the bare names by family, not as strings, but OpenAI's model pages mark gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini as "Deprecated," and each one's default snapshot is on the Jan. 14 list (GPT-4o, GPT-4o mini, GPT-4.1, GPT-4.1 mini). Treat an unpinned gpt-4o as affected. And gpt-4o-2024-05-13 isn't on the list because it's already gone, shut down Oct. 23, 2026 (OpenAI deprecations). If something still calls it, it has been failing for 20 days.
The page covers the API only. It says nothing about ChatGPT or Azure OpenAI, and OpenAI's pricing page notes that its models on Microsoft Azure "are billed through those services" (OpenAI pricing). If you run on Azure, check Microsoft's schedule instead of assuming this date.
This is not the Dec. 11 deadline
Different cutoff, different product. Custom GPTs inside ChatGPT "are scheduled to retire on Dec 11, 2026," with Feb. 11, 2027 for FedRAMP workspaces and approved Enterprise deferrals (OpenAI Help Center). That's GPTs built in the ChatGPT app moving to plugins. Jan. 14 is about code and automations calling the API.
To keep things interesting, Dec. 11 is also an API date. Older gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 and o3-pro snapshots shut down that day (OpenAI deprecations). If someone "upgraded" your flow to gpt-5-mini last year, that's the earlier fire.
Where the model string lives
Zapier. In a ChatGPT (OpenAI) step, Zapier's help docs tell you to "click the Model field to expand the dropdown menu" (Zapier). Its action comparison page, last updated April 6, 2026, still lists "GPT-4o GPT 4.1 GPT 4o mini" among the Conversation action's supported models (Zapier). If you use AI by Zapier without your own key, you pick from the same kind of dropdown and pay in tasks: GPT-4o costs 5 tasks a step and GPT-6 Luna costs 1, per Zapier's September 2026 list, which still includes all four affected models (Zapier). As of Nov. 12, we found no Zapier notice about Jan. 14, so open the step and change it yourself.
Make. In the OpenAI (ChatGPT, Whisper) app, the Generate a completion module's Model field says "Select or map the model you want to use" (Make). "Map" is the trap. The model name can come from an earlier module, like a spreadsheet cell, so check the source too. Make's setup page still suggests switching to "a non-reasoning model (GPT-4o, or GPT-4.1)" if organization verification blocks reasoning models (Make). That's advice with an expiration date.
n8n. In the OpenAI node, the Text resource offers Generate a Chat Completion and Generate a Model Response, each with a Model parameter. For the first, n8n's docs still say "try gpt-4o if you need high intelligence or gpt-4o-mini if you need the fastest speed and lowest cost" (n8n Docs). Plenty of workflows were built by people who followed that line.
Custom code. Search the code and environment variables for gpt-4o and gpt-4.1. OpenAI's List models endpoint returns a shutdown_date field for exactly this (OpenAI deprecations), and Make's Make an API call module can query /v1/models without any code (Make).
Defaults and aliases won't save you
A bare alias isn't a hedge. gpt-4o points to the default snapshot gpt-4o-2024-08-06 (OpenAI), which is on the list. The last cleanup is the precedent: when gpt-4-0613 shut down Oct. 23, the gpt-4 alias was listed right beside it (OpenAI deprecations).
When a tool picks the model for you, the vendor decides when it changes, and vendor docs lag. Zapier's comparison page still lists babbage-002, davinci-002 and gpt-3.5-turbo-instruct for Send Prompt (Zapier). OpenAI shut all three down Sept. 28, 2026 (OpenAI deprecations). When the Assistants API went, Zapier's notice was blunt: "On August 26, 2026, Zaps using these actions will stop working" (Zapier).
Our favorite detail: when OpenAI retired two GPT-4 snapshots on March 26, 2026, its recommended replacement was gpt-5 or gpt-4.1 (OpenAI deprecations). The replacement now needs a replacement.
Test Luna on real tickets, not a demo prompt
Pull 20 to 50 real tickets, quote requests or emails from the last month. Run them through your current model and gpt-6-luna side by side, and have whoever normally checks the output grade both without knowing which is which. Do it before the holidays eat December.
Watch the settings. Luna is a reasoning model with medium as its default effort, and in Chat Completions it supports function calling only with reasoning effort set to none (OpenAI). Make says the same thing: None is "Required to use Tools (functions) with GPT-5.6 and GPT-6 Sol or Luna models" (Make). And reasoning tokens are "billed as output tokens" (OpenAI reasoning guide), which changes the cost math.
The cost math
This is a hypothetical workload, and the math is illustrative: a support reply flow handling 2,000 tickets a month at 1,500 input tokens and 300 output tokens per ticket. Prices are standard list rates per million tokens from OpenAI's pricing page.
gpt-4oat $2.50 in, $10.00 out: (1,500 × $2.50 + 300 × $10.00) ÷ 1,000,000 = $0.00675 a run. Times 2,000 is $13.50 a month.gpt-4.1at $2.00 and $8.00: (1,500 × $2.00 + 300 × $8.00) ÷ 1,000,000 = $0.0054 a run, or $10.80 a month.gpt-4.1-miniat $0.40 and $1.60: (1,500 × $0.40 + 300 × $1.60) ÷ 1,000,000 = $0.00108 a run, or $2.16 a month.gpt-4o-miniat $0.15 and $0.60: (1,500 × $0.15 + 300 × $0.60) ÷ 1,000,000 = $0.000405 a run, or $0.81 a month.gpt-6-lunaat $0.10 and $0.50, excluding reasoning tokens: (1,500 × $0.10 + 300 × $0.50) ÷ 1,000,000 = $0.0003 a run, or $0.60 a month.
Coming from GPT-4o, that's ($13.50 minus $0.60) ÷ $13.50 = 95.6% cheaper on paper. Coming from GPT-4o mini, the edge is thin, and reasoning can erase it. Luna's input costs $0.00015 a run. To stay under GPT-4o mini's $0.000405, output can cost at most $0.000255, which at $0.50 per million is 510 tokens. Take out the 300 visible tokens and you have 210 reasoning tokens to spare. OpenAI says reasoning models "may generate anywhere from a few hundred to tens of thousands of reasoning tokens" (OpenAI reasoning guide). Log the real count during your test, then try effort none.
At this volume the bill isn't the risk. A flow that quietly errors on Jan. 14 is. If usage pricing has burned you before, our surprise AI bill guide covers caps. Zapier tasks are a separate meter, which we priced out in what 5,000 automations cost.
Not sure which model your flows call? Send us the Zap, scenario or workflow and we'll find the strings: team@lirium.ai.