It was early evening, and I was working on a prototype of my wallpaper app. All I asked for was one short line of helper text under the Save button on the detail screen. What came back added the line, and then also reworked the button's padding and resized the icon.
That kind of request usually ends with exactly one new line. My first thought was that the assistant was being unusually thoughtful. My second thought was that I couldn't tell whether it had been thoughtful or had simply wandered off scope.
Only later did I notice the model menu. It said GPT-6.2 Astra. I was fairly sure I had picked GPT-6 Astra.
What changed: the November 18 changelog entry
Rork's changelog (rork.com/changelog) lists the following under November 18:
- GPT-6.2 Astra replaces GPT-6 Astra, and GPT-6.2 Sol replaces GPT-6.1 Sol
- Projects that had the older version selected now build with the new one at the same effort level
- Both are included in the Pro and Max plans
The same entry also removes the 20-site limit on websites you can publish from one project, and fixes App Review submissions failing while an in-app event was still waiting for review. Unrelated to models, but worth a glance if you're about to ship.
In short, the model doing the building had moved up a version without me touching anything.
The mistake I almost made: rewriting my prompt
When I saw the extra padding changes, I assumed my request had been vague. I had already typed "Add one line of helper text. Don't change anything else." and was about to send it.
Then I saw the menu and stopped. If the model change was behind the different answer, no amount of prompt polishing would get to the cause. And jumping to "the new model is worse, rebuild the screen" would have been just as premature.
When something feels different, count what actually changed before fixing anything. These days I run three checks, in this order.
Check 1: the model name and effort in the menu
First, look at the menu itself:
- Does the model name now say "6.2"?
- Is the effort still the level you chose earlier?
If the changelog is right, effort should not have moved. If it did, that wasn't the automatic switch. Someone else on the project, or a past version of you, changed it. On a shared project that's worth confirming.
I wrote about how I pick an effort level in the article where I built the same grid screen with Opus 5.5 and GPT-6 Sol. Whether to raise it can wait until this check is done.
Check 2: mark the last version built with the old model
Next, note which version was the last one the old model produced. Rork saves the whole project as a version after every answer, and Undo or Revert cost no credits. I covered how to choose between them in the Undo, Revert, and Edit message article.
If you export the project to GitHub, a tag on the last pre-switch commit makes later comparisons easy:
# Mark the last state before the switch (pick the commit from your own history)
git tag before-gpt-6-2 <commit-before-switch>
git push origin before-gpt-6-2
# After one request on the new model, count how many files moved
git diff --stat before-gpt-6-2 HEADI use git diff --stat because I want to compare answers by how many files and lines moved, not by gut feeling. For my detail-screen request, one line of text had touched 3 files. Once it's a number, deciding between "thoughtful" and "out of scope" gets much calmer.
Check 3: send the same yardstick request once
The third check is a small request you always send with the same wording. Mine is "Make the section headings on the Settings screen one size smaller." The answer is short, and I know exactly one file should change.
After the switch, I send it once and look at two things:
- Did the change stay within that one file?
- Did it touch anything I didn't ask about, like spacing, colors, or ordering?
If the yardstick shows no difference, what happened on the detail screen was probably about how I phrased the request. If it does show a difference, I note that the new model tends to reach wider, and start adding "limit the change to this one file" to my requests. Either way, there's no need to rebuild the screen.
Don't confuse it with the other model
One thing is easy to mix up. What changed here is the model Rork uses to build your app. The AI model your app calls at runtime, the model name written in your app's code, didn't change by a single character.
I'll admit I first assumed both had switched, and went digging through my app's code. App-side model names have their own deadlines. Claude Sonnet 4.5 retires on November 30, for example, and that only gets handled if you edit the code yourself. I wrote up how to find those IDs in the article on scanning your app for old Claude model IDs.
Keep the model that builds your app and the model your app calls in separate notebooks. As an indie developer running several apps side by side, I can't hold both in my head, so I split them right at the heading level of my notes.
What to do next
Pick one yardstick request and send it once in a project that has switched. Counting the files it moves usually tells you whether a rebuild is needed at all.
Compared with the days when every model change made me scramble, running these three checks now feels like setting something heavy down for a moment.