I’m currently experiencing an issue where every interaction with Cappy ends in a server error.
Based on my previous experience, this usually happens when the credit budget is depleted. However, in the top-left corner, it still shows a remaining budget of 6,984.22 credits. Interestingly, the bar below the number appears to be empty.
I’m now wondering which information is correct and whether adding more credits would solve the issue.
Hi @leosch ,
Thanks for the details.
Could you try switching model to GPT-5.5 in the chat and testing again?
Thanks for your quick reply!
I’ve already tried several models across different projects and bases, including GPT-5.5, but the error remains the same.
Hi @leosch ,
Please try using the models under the Recommended section first, such as GPT 5.5 or GPT 5.4 Mini. These are the models provided by Teable.
We recently refactored Agent Computer, so BYOK needs to be made compatible again. BYOK is not supported in the new Agent Computer yet.
For now, please use the Recommended models and test again.
That actually worked. I used GPT 5.5 from the Recommended section and got at least some kind of answer. However, I’m now getting a different error. Could this one possibly be related to the credit issue?
Hi @leosch ,
This looks more like a temporary error.
Could you click Retry and test again? If the same error keeps appearing, please send us the App ID, and we’ll continue checking.
Thanks for the support, I got the system back up and running.
The problem is that ChatGPT 5.5 is far from useful for our use case. Claude was much more effective, had a significantly better understanding of what we wanted from it, worked much faster, and basically outperformed ChatGPT in every area.
Some tasks even seem almost impossible with GPT compared to Claude. Is there already a plan for when Claude will be available again?
Hi,
Thanks for the honest feedback.
We updated the model selection because GPT-5.5 currently gives us a better balance of overall intelligence, cost efficiency, and user experience. This also helps us let users use Teable more freely without worrying too much about credits.
That said, we understand that Claude worked better for some specific workflows. We’ll keep evaluating model options, but we don’t plan to change the current model setup in the short term.
To help us assess this properly, could you share a few concrete examples where Claude performed much better than GPT-5.5? Specific real use cases would be very helpful for us.
I wanted to set the right expectation upfront so it doesn’t affect how you plan to use the product.
Hi,
attached, you will find a screenshot of a task for Cuppy where objects within an application were supposed to be made draggable. GPT-5.5 took a full 41 minutes for what, in my view, was a fairly simple task. In that amount of time, Claude had already completed an entire app.
For comparison, a request to Cuppy using Claude to implement a completely new saving approach took less than five minutes.
On top of that, there were major errors and user-unfriendly adjustments that Claude never made. In the following example, GPT was simply supposed to add a button that allows a task in the app to be marked as mandatory or optional. This was a simple instruction, but GPT was not able to follow it properly. Instead, it added a dropdown menu with options that had nothing to do with the task. So the request was clearly misunderstood and the AI hallucinated the implementation. Nevertheless, this simple request took 27 minutes, only for the result to be unsatisfactory.
In this context, it would be important for us that our Claude API key, or the BYOK function, becomes available again soon so that we can return to a normal working rhythm.
Hi @leosch ,
Thanks for sharing the detailed comparison.
We understand your concern. The longer execution time and the implementation quality you described are both valuable feedback for us.
BYOK has now been restored, so you can try using your Claude API key again.
We’ll also review GPT-5.5’s performance in App Builder and see where we can further improve the experience.
Thanks again for taking the time to document these examples.
