I am starting to rip my own hair out on this issue. I get the same error message “An internal error has occurred” all the time. It started out appearing randomly, the length of the message doesn’t seem to be the issue. The Token size of the entire conversation is barely 40k tokens. But now the error happens in almost every message.
And it is especially frustrating that there is absolutely no info about this on the documentation, and all I found when researching this issue was a lot of people mentioning a similar problem, and never getting answers either.
Maybe it is a very “on the nose” thing, I would still be happy if someone just pointed it out to me. I can not figure it out by myself.
I am using Firefox, but using other browsers doesn’t change the issue.
I tried changing the models, that doesn’t make a difference.
I tried downloading the chat history and re-uploading it in a new chat, but somehow that doesn’t work either.
I tried creating a completely new conversation and re-inserting all relevant information for my project, but the answers from the model are so vastly different even though the information is the exact same.
I thought that maybe the servers were overloaded, I wanted to change to the billed API plan, but for some reason I can’t get the billing to work. So now I am at a standstill.
If anyone has any idea about what could be causing this, please let me know.
You can switch on 9 different charts to explore your requests, and you can look at the status codeas.(200 ok, 500 internal server error, 429 limit exceed)
In the last 6 hours, I’ve made 99 requests and 6 were successful.
Request no. 7 gave me an 500 server error so I waited and retried. Each time it retried I got another error 500 until I hit the “50 requests per day limit” and I now get 429 errors for hitting the limits.
Grumpy mode:
It’s kind of Google to count requests that fail due to Google’s ownl server errors towards our rate limits!
Why is it whenever Google release new models/functionality it breaks existing models already in production?
Either wait and try again or try switching model as they tend to be broken at different times. I’m using model Gemini-1.5-Pro-001
I keep getting the same issues over the past several days, for saved things I’ve ran many times. It was only erroring if I chose 1.5 Pro 002 but now it does it on anything I choose
Getting the same error almost all the time, Google please do something, I am on the paid plan and getting this instead of prompt responses is not cool!
Getting smacked with this a lot. Particularly after I’ve gotten a rather large response, if I ask a question afterwards I’ll get the error, but if I clear the large response and re-do it I can continue sans errors.
Tambien estoy presentando en el mismo probrlmas de mucho aqui “Se ha producido un error interno”
Me puedes informar como hago para pagar un plan premium de la IA “PENSAMIENTO EXPERIMENTAL GERMINI 2.0 FLASH 01_21”
iI switched to Flash 2.0, and everything started working. It’s a shame that I can’t use more advanced models, but at least I can use it somehow, so thanks for that.
(Friends, let’s not be pushy, it’s free after all!)
Hi everyone,
I wanted to share some insights based on recurring issues, along with the solutions that worked for me. Hopefully, this helps others dealing with similar problems.
Token Limit Inconsistencies
One of the most frequent issues I encountered was related to token limits. Surprisingly, even when my inputs didn’t seem close to the model’s token max, converting my files to plain .txt revealed that the actual token count exceeded the limit. This showed me that token extraction can behave inconsistently depending on the format—some encodings or hidden characters may inflate token count beyond what you’d expect. Recommendation: Always check token usage with plain text conversions before assuming you’re within limits.
Variability in Model Responses
It’s completely normal for the model to return significantly different outputs, even when using the same prompt, especially under non-deterministic settings.
If you want consistent responses:
Adjust temperature and top-p to lower randomness.
Or, for exact replication: manually copy the previous response into the editor and prompt the model to continue or regenerate based on that.
**“An internal error has occurred”
This error appeared most often during extended conversations. I’m using the Brave browser, and I had Memory Saver enabled. In long sessions, a single AI Studio tab could consume over 2 GB of RAM, and with Memory Saver on, the tab wasn’t prioritized for RAM allocation—leading to internal crashes.
Fix:
I disabled Memory Saver entirely and allowed full RAM usage for the AI Studio tab.
Additionally, I now use Google AI Studio exclusively on a desktop machine where RAM resources are more stable and not restricted by browser-level optimization.
Most of these issues stemmed from local setup or browser behavior, not from the model or Google’s infrastructure itself.
I’m on a pretty powerful desktop machine (I’m a game developer, so it’s a must) but Google AI Studio & Gemini 2.5 Pro Experimental have suddenly started “forgetting” huge portions of the conversation, even though it’s supposed to be saving to my Google Drive (which has ample free space) and now won’t respond at all other than to tell me that “An internal error has occurred”… I’ve killed all other processes, tried several browsers, and done most anything else I could think to, but to no avail. I was also having issues with it accepting image input recently; the “loading” icon would just spin on infinitely. I absolutely love the model, when it works… but this is absolutely obnoxious. I’ve tried several things including switching to older models, but the results are the same. It seems Google AI Studio just “breaks”. Terribly frustrating as we were making good progress on my current project. I’m left with no choice but to summarize the existing data and start again with a fresh conversation… one that has no other context, and well, that’s far from ideal. In short: Grrrrrr…
Interesting… I hadn’t thought to clear out the former responses. Deleting that information is easy enough, I suppose, though it is a pity to rob the AI of that contextual content. Still, worth a shot if I can continue working without having to essentially start from scratch. Thanks for the suggestion!