The AI Your Staff Uses Changed Overnight
GPT-6 is now the default in ChatGPT. Here is a ten-minute routine to catch what changed before a customer does.
By John Martines · October 9, 2026 · 6 min
The AI tools your team relies on get new models without asking you. A short "golden set" of your real tasks, rerun after every change, tells you what's different before it reaches a customer.
Your office manager opens ChatGPT on Wednesday morning and pastes in the same prompt she has used every week since spring: turn this list of open invoices into a polite reminder email. It works. But the email is longer than last week's, the tone is a little warmer, and the totals now come back in a neat table with a button next to it. She copies it into Outlook and the table falls apart.
Nothing is wrong with her prompt. The AI underneath it changed overnight.
A "model" is the engine inside an AI tool. ChatGPT, Copilot, Gemini and Claude are the apps your staff see. The model is the part that actually writes the answer, and vendors swap it out regularly, usually for a better one. The app looks the same. The answers don't always.
This week's swap is a big one:
- September 22, 2026: OpenAI introduced GPT-6 Sol and GPT-6 Luna, first in Codex and ChatGPT Work for paid plans, and in the desktop app for Free and Go users.
- October 7, 2026: GPT-6 became the default model in ChatGPT's chat for Plus, Pro, Business and Enterprise accounts, powered by GPT-6 Sol. It arrived with "Intelligent UI," which lets an answer include buttons, forms, charts, diagrams and small working calculators instead of plain text.
- October 8, 2026: Free and Go accounts started getting GPT-6 Luna as their default.
No one had to turn anything on. If your team uses ChatGPT, they are using a different model this week than last week.
It's not the first time, and it won't be the last:
- April 2025: OpenAI released an update to GPT-4o that made ChatGPT overly flattering and agreeable. It rolled the update back about four days later and published an explanation on April 29, 2025.
- August 2025: GPT-5 launched and replaced the older models in ChatGPT. Enough people missed GPT-4o's style that OpenAI brought it back for paying users behind a "show legacy models" setting.
- February 13, 2026: OpenAI retired GPT-4o, GPT-4.1, GPT-4.1 mini and o4-mini from ChatGPT for good, saying only 0.1% of users still picked GPT-4o on a given day.
And it's not just ChatGPT. Microsoft made GPT-5.6 the preferred model across Microsoft 365 Copilot in July 2026, set Claude Sonnet 5 as the default for higher-reasoning work in Word in August, and on October 6 added a "Switch Model" option to Copilot's regenerate button. Google made Gemini 3 Pro available in the Gemini app for Workspace customers in November 2025, rolling out automatically over about two weeks. Anthropic retires older Claude models on a published schedule; Claude Sonnet 4.5 retires from the Claude API on November 30, 2026.
Most of the time, a new model is better. GPT-6 likely writes cleaner drafts than what your team had in the spring. The issue isn't quality. It's that your business built habits on top of the old one.
Think about what your staff have quietly come to rely on. A prompt that turns job notes into a customer summary. A checklist that asks ChatGPT to review a quote. A saved template that writes the weekly social post. Each of those was tuned, by trial and error, to how one model behaved. When the model changes, the answer can change in ways that are easy to miss: a number that used to be right, a step that's now skipped, a tone that's friendlier than you want with a late-paying client, or a nice interactive chart that doesn't survive being pasted into an email.
None of that is a reason to worry. It is a reason to look.
1. Write down your golden set. Pick the three to five AI tasks your team does most often, the ones whose results go to a customer, a client file or your books. For each one, save the exact prompt, the input you give it (a sample invoice list, a sample job note) and a copy of an answer you were happy with. Put it in one shared document. That's your golden set. It takes about half an hour once.
2. Rerun it when the vendor announces a change, and once a month anyway. This week is the first rerun. After that, put a monthly reminder on your AI champion's calendar. Many changes are announced only in release notes, and some arrive gradually, so the monthly check catches the quiet ones.
A free account opens the rest of this article and the whole library.