ChatGPT switched to GPT-6. How do you check your AI still works?

Model updates are usually good news. A short re-test tells you whether they are good news for your work.

Check AI after model update brief

0:00

/90.557823

Your AI just got an upgrade

On 7 October, OpenAI started rolling out GPT-6 in ChatGPT. Paid plans got it first, and free users followed from 8 October. For most people the switch happened automatically. According to OpenAI's help centre, GPT-6 replaces "Latest" in the model menu, and anyone who had "Latest" selected was moved across automatically.

That is a smooth way to ship an upgrade. OpenAI says GPT-6 starts answering questions that need a web search 44% sooner on average, and it can now reply with charts, buttons and small interactive tools as well as text. Plenty of people will notice nothing except that answers feel quicker and look better.

But think about the AI work your business now relies on. The prompt that drafts your quotes. The template that summarises customer calls. The routine that turns a messy email thread into a task list. Those were all tuned, a little at a time, against the model you had last week.

This is a great week to give them a quick tune-up.

Every AI tool does this, and that is good news

Every serious AI product updates the model underneath it, often by default, and often without changing the name on the button. Google's new Gemini agent, announced on 8 October, picks the best model for each task by default, and users can choose a different one if they want. It is the same pattern: the tool decides which engine runs your work, and the engine keeps improving.

Even the same model name can mean different things in different places. OpenAI's own safety notes for this release explain that ChatGPT's Chat tab now uses the October versions of GPT-6 Sol and Luna, while Codex and ChatGPT Work are still on the September versions.

Newer models are usually better. A quick re-test is how you turn "better on average" into "better for your quote template". An upgrade can change things you never thought to write down:

Length and tone. Answers might get shorter, longer, warmer or more formal.
Format. A reply that used to be plain text might now arrive as a table or an interactive card that does not paste neatly into your CRM.
What it looks up. A model that searches the web more readily might pull in fresher information, so check it still uses the background you gave it.
How literally it follows instructions. A rule that was always respected might now be read a little differently.

Each of these is small. Spotting them early is what keeps a good workflow getting better with every upgrade.

A 30-minute re-test kit

You do not need a testing department. You need a small set of examples and someone who owns them.

Pick your five to ten most important AI jobs. Choose the ones that touch customers, money or decisions, not the ones you use for fun.
Save a "golden" example for each. Keep the input you gave it and an output you were happy with. A shared folder is enough.
When the model changes, run them again. Paste the same inputs and put the new outputs next to the old ones.
Check what matters, not every word. Are the facts right? Is the format still usable? Did it follow the rules you care about? Would you send it?
Write down the result. Note the date, who checked, and anything that changed. If something works differently, tweak the prompt, and keep the improvements you like.

The first round takes about half an hour. After that it is quicker, and you will start to spot which jobs are sensitive to change and which never move.

Know what you are actually running

A simple habit makes upgrades easy: know which AI tool, model and setting each important workflow uses. Write it down. One line per workflow is plenty.

Where a tool lets you pin a model, consider doing it for your most sensitive jobs. ChatGPT, for example, keeps people who had explicitly chosen GPT-5.6 Sol on GPT-5.6. That gives you time to test before you move. Then keep an eye on the places vendors announce changes, such as release notes, help centre articles and changelogs. OpenAI's own developer changelog has, in the past few weeks, advised developers to "rerun your evaluations" after a fix that affected image understanding. That is good advice for everyone, not just developers.

Give the signal to one person

The kit only works if someone runs it. Name one person who gets the "the model changed" signal and owns the re-test. Treat it like a supplier changing the spec of a part you build into your product. You would welcome the better part, and check it fits before the next order went out.

Where Walter fits

At Walter, we are model-agnostic on purpose. The best model for a job will keep changing, and that is a good thing. What should not change every time a model does is your business context: how you price, who approves what, what your customers expect, and what good work looks like.

Walter is being built to keep that context together in one place, across whichever AI models you use, so every model upgrade is something your team can test, adopt and benefit from quickly.

So this week, while everyone is enjoying the new answers, spend thirty minutes on your golden examples. Upgrades are good news. Checking is how you make sure they are good news for you.