Check The Expiration Dates

Every AI tool in your business has a carton date stamped on it. Most owners have never looked.

Last Friday a pile of older OpenAI models quietly stopped answering: GPT-3.5 Turbo, GPT-4, GPT-4 Turbo, the original gpt-4o snapshot, o1, o3-mini and a handful of others. If something in your shop was still pointed at one of them, it broke. Maybe loudly, or maybe in a way nobody's caught yet, like a chat widget that now just spins.

And that was the warm up. OpenAI's notice dated October 14 pulls GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini out of the API on January 14. Those were the default models behind almost every chatbot, review responder and lead qualifier an agency sold anybody in 2024 and 2025. Meanwhile Google is turning Gems into Skills and Notion is changing how its AI credits work, so the dates keep stacking up between now and spring.

The part that should bother a small business owner is that you probably don't hold the API key. Your freelancer does, or your agency, or your chatbot vendor, or some Make scenario a contractor built two years ago and never mentioned. The shutdown email went to them. You're trusting they read it.

So this week is an expiration audit. You'll spend an hour finding what's running on borrowed time, send one email to everyone holding a key, and set up a test routine so the swap doesn't blow up in January.

Open a calendar or a sheet you'll actually look at, not the project board you abandoned in March, and put these dates in it.

October 13. You can no longer create or edit Gems in Gemini. That one's already behind us. Android Central has the timeline.

October 23. Also passed. The original gpt-4o-2024-05-13, gpt-4.1-nano, GPT-3.5 Turbo, GPT-4, GPT-4 Turbo, o1, o1-pro, o3-mini, o4-mini and gpt-image-1 are gone. OpenAI pointed people to gpt-5.6-sol, gpt-5.6-terra or gpt-5.6-luna as replacements, and gpt-image-2.5 for images.

November 2. Notion moves existing customers to its new AI credit system. Check what your plan actually gets before the bill tells you.

November 17. Personal account Gems migrate automatically to Skills. Opal, Google's mini app builder inside Gemini, also shuts down in November for personal accounts.

November 30. OpenAI's Evals platform, Agent Builder and reusable prompts go away.

December 1. GPT Image models.

December 11. Older GPT-5 and o3 snapshots, including gpt-5-2025-08-07 and gpt-5-mini-2025-08-07.

January 6. tts-1, tts-1-hd and gpt-4o-mini-tts move to gpt-realtime-2.1-mini. Same day, you can't start new fine tuning jobs on the GPT-4o and GPT-4.1 bases.

January 14. The big one. gpt-4o-2024-08-06, gpt-4o-2024-11-20, gpt-4o-mini-2024-07-18, gpt-4.1-2025-04-14, gpt-4.1-mini-2025-04-14 and every fine tune built on them. Plain aliases like "gpt-4o" point at these snapshots, so treat the aliases as dead too. OpenAI recommends gpt-6-luna across the board.

January 20. Legacy audio and realtime models, including gpt-4o-audio and gpt-4o-realtime. If you've got an AI phone receptionist, circle this one twice.

February 26. whisper-1, gpt-4o-transcribe and gpt-4o-mini-transcribe. Anything that turns calls or voicemails into text lives here.

March 2027. Gems disappear for Workspace business, enterprise and nonprofit accounts, per Google's admin page. Education gets until June.

April 1. gpt-5.3-codex, gpt-5.1 and gpt-5.4-nano.

One more note for the calendar. OpenAI's stated policy is at least six months notice on generally available models, and the January notice gave three. Don't hold your breath for an extension.

Set a timer. You're looking for model names, and they hide in the same few places.

Make scenarios. Open Make.com and go scenario by scenario. Every OpenAI module has a model field. If it says gpt-4o, gpt-4o-mini, gpt-4.1 or gpt-4.1-mini, write it down next to the scenario name. Check HTTP modules too. Plenty of builders skip the OpenAI module and call the API direct. Then the model name sits in the request body.

Go High Level. In Go High Level, look at workflows with AI steps and any custom webhook actions. Webhooks are where freelancers like to wire in their own OpenAI calls, so open each one and read the payload.

Your website chatbot. If a vendor runs it, you won't see the model from your side. Write down the vendor and whoever you actually talk to there. The email below handles the rest.

Old custom GPTs and Gems. Go through ChatGPT and Gemini and list every custom GPT and Gem your team leans on. Gems have been frozen since October 13, so figure out which ones matter before they turn into Skills.

Freelancer scripts. That Python script that cleans your lead list every night. The Google Apps Script bolted onto your quote sheet. Ask whoever wrote them, or open the file and search for "gpt".

Env files and secrets. If you have any code at all, have someone search the code, env files and stored secrets for "gpt-4o" and "gpt-4.1". For a developer it's a five minute job.

The usage dashboard. If you own an OpenAI account, the usage page shows which models are actually getting traffic. A developer can also hit the List models endpoint. It returns a shutdown_date for anything on the way out.

About twenty minutes in, you'll hit a wall: a Make export or a vendor reply so thick with jargon you can't tell whether there's a model name buried in it. Paste it into Claude or ChatGPT with this:

You don't need to understand their system. You need model names and dates, in writing. Send this to every agency, freelancer and vendor on your list.

Question five is the one I care about most. Anybody who answers with "we'll just switch it over" hasn't thought about it yet, and you'd rather learn that in October.

For anything on the January list, OpenAI points you at gpt-6-luna. For the stuff that died Friday, the suggestions were the gpt-5.6 family. Your vendor may have a good reason to pick something else. Fine, as long as they tell you what and why.

Then test it properly. This is the routine I'd hold any vendor to:

The replacement is a lot cheaper. Per million tokens, input and output:

GPT-4o: $2.50 in, $10.00 out.

GPT-4.1: $2.00 in, $8.00 out.

GPT-4.1 mini: $0.40 in, $1.60 out.

GPT-4o mini: $0.15 in, $0.60 out.

gpt-6-luna: $0.10 in, $0.50 out.

Say you push 10 million input tokens and 2 million output tokens a month. That's about $45 on GPT-4o and about $2 on gpt-6-luna. On GPT-4o mini the savings are real but small. The big wins are anything still sitting on GPT-4o or GPT-4.1.

Cheap matters more than it used to. AI Weekly reported that legal AI company Harvey watched its margins go from about 50% to minus 50% as token usage grew 20x.

Now the trap. gpt-6-luna is a reasoning model, meaning it thinks before it answers, and those thinking tokens get billed as output. So your token count can jump while your dollar cost drops. If you or your vendor set alerts on token usage, they'll start screaming the week you switch, and somebody will panic and roll back a perfectly good migration.

Head that off by changing every usage alert to a dollar threshold before you switch. I'd pick a number a little above your current monthly spend and alert on that.

Forward this part as is. They'll know what it means.

Function calling. gpt-6-luna defaults to medium reasoning effort. On the Chat Completions API, function calling only works with reasoning effort set to "none." If your setup needs tools and reasoning together, it has to move to the Responses API. Miss this and anything that books, looks up or tags will fail quietly.

Output limits. GPT-4o capped output at 16,384 tokens, and plenty of code has its own truncation tuned to that number. gpt-6-luna can return up to 128,000. If the client side still chops at the old limit, long outputs get cut off mid sentence, so check for hardcoded limits.

Time to first token. Reasoning adds a pause before the first word appears. In a background job nobody cares. In a website chat widget, a visitor staring at an empty bubble for several seconds is going to leave. Test the widget in a browser like a customer would. If it drags, try a lower reasoning effort and measure again.

Ask about fine tunes too. If you paid for a fine tuned GPT-4o or GPT-4.1 model, it stops on January 14 along with the base. New fine tuning jobs on those bases are blocked from January 6, so any plan to retrain on the old base has a hard deadline.

Picture a three dentist office. In 2024 their agency built two things: a chat widget on the website that answers insurance questions and books cleanings, and an automation that drafts replies to Google reviews. Both run on gpt-4o-mini. The office manager has no idea, because the agency holds the key and the invoice just says "AI chatbot, monthly."

She does the hunt one Tuesday afternoon between patients. Their own Make account is clean. There's one GHL workflow with a webhook pointing at the agency's server. She sends the email.

Three days later the agency writes back with a paragraph about "our proprietary AI stack." She pastes it into the prompt above, which digs one model name out of the second paragraph (gpt-4o-mini) and points out that the reply never answered the rollback question.

She writes back asking for the test plan. The agency runs 50 old chats and 50 old reviews through gpt-6-luna. What came back:

Review replies. Tone is fine, but about a third came out longer than the old ones. They tighten the prompt to cap replies at three sentences.

Chat widget booking. Broke on the first run, because the booking step uses function calling on Chat Completions with default reasoning. They set effort to none for the booking flow and it works again.

Chat widget speed. With reasoning on, the first word took long enough that the widget felt broken. Booking is already on effort none, so they try a lower setting for the insurance questions until it feels snappy.

Billing alerts. The agency's token alert would have tripped in week one. They switch it to dollars.

The savings are modest, since gpt-4o-mini was already cheap. The real win is a chat widget that doesn't go dark on a Thursday in January with a waiting room full of people. They go live in December and keep the old setup as a fallback until January 13.

This won't be the last batch. April already has three OpenAI models on the list, and Google's Workspace Gem cutoff lands in March. Put a 30 minute block on the first Monday of every quarter.

When it comes up, open the expiration calendar and delete what's passed. Check the providers' deprecation pages for anything new, rerun the hunt on whatever automations got built since last time, and resend the vendor email to anyone who's changed something. Then check your AI spend in dollars.

Thirty minutes, four times a year, beats one Saturday spent figuring out why the chatbot stopped booking appointments.

Put every date above on a calendar you actually check.

Spend one hour on the hunt. Make, GHL, chatbot, custom GPTs, Gems, scripts.

Send the vendor email to everyone who holds a key you don't.

Run any confusing reply through the prompt and match the model names to your calendar.

Switch every AI usage alert from tokens to dollars.

Book your first quarterly check for January.

Jordan

The AI Newsroom | Practical AI for people with a business to run.