I’m on a paid key.
I just got it to run by simply appending a full stop to the prompt after each unsuccessful attempt. Seems like it only keeps failing if the same prompt is repeated.
I believe that adding a small, unique identifier will help a lot after my many experiments! You can try that way.
I tried this didn’t help at all. Are we fooling AI to ignore such basic things?
it wont help this morning it was working fine for like 1:30 hour then it broke again
2 Likes
TPM has been 250k per minute for a while. The service is broken. Either they are overloaded with lot of requests/some sort of upgrading is happening.
It works flawlessly for me since a couple of hours. Let’s hope they fixed it finally.
EDIT: ~70 requests without a single empty response and pretty good latency
1 Like
Its definitely not resolved 
we have a scenario where we specifically handle this case and it was happening 2-3 hours back.
yeah it was fixed now itss broken again
Having the same problem with 2.5 pro today, it used to be infrequent but today it’s almost every request. Not having any issues with flash strangely.
1 Like
almost 2/3 requests fail. facing these issues the whole week already…
unusable at current point
2 Likes
quá nhiều yêu cầu bị mất phần text trong phản hồi, tốn token (my money) và thời gian, nhưng tôi lại nhận về một kết quả rỗng, một trải nghiệm cực tệ, khi chuyển qua 2.5 Flash thì thành công 100%, nhưng model Flash không đáp ứng được yêu cầu câu trả lời cho tôi. đây là logs
[DEBUG GEMINI:gemini-2.5-pro] Raw response (len=200): {
0|bot | “error”: {
0|bot | “code”: 500,
0|bot | “message”: “An internal error has occurred. Please retry or report in Troubleshooting guide | Gemini API | Google AI for Developers”,
0|bot | “status”: “INTERNAL”
0|bot | }
0|bot | }
0|bot |
0|bot | [DEBUG GEMINI:gemini-2.5-pro] Status: 500, OK: false
0|bot | [GROUP CHAT GEMINI] (gemini-2.5-pro) Lần thử 2/5…
0|bot | [DEBUG GEMINI:gemini-2.5-pro] Raw response (len=614): {
0|bot | “candidates”: [
0|bot | {
0|bot | “content”: {
0|bot | “role”: “model”
0|bot | },
0|bot | “finishReason”: “STOP”,
0|bot | “index”: 0,
0|bot | “groundingMetadata”: {}
0|bot | }
0|bot | ],
0|bot | “usageMetadata”: {
0|bot | “promptTokenCount”: 5366,
0|bot | “totalTokenCount”: 5580,
0|bot | “promptTokensDetails”: [
0|bot | {
0|bot | “modality”: “TEXT”,
0|bot | “tokenCount”: 1236
0|bot | },
0|bot | {
0|bot | “modality”: “VIDEO”,
0|bot | “tokenCount”: 3682
0|bot | },
0|bot | {
0|bot | “modality”: “AUDIO”,
0|bot | “tokenCount”: 448
0|bot | }
0|bot | ],
0|bot | “thoughtsTokenCount”: 214
0|bot | },
0|bot | “modelVersion”: “gemini-2.5-pro”,
0|bot | “responseId”: “6VWfaJPaCNuez7IP2ZK9mAg”
0|bot | }
0|bot |
0|bot | [DEBUG GEMINI:gemini-2.5-pro] Status: 200, OK: true
0|bot | [GROUP CHAT GEMINI] (gemini-2.5-pro) Lần thử 3/5…
0|bot | [DEBUG GEMINI:gemini-2.5-pro] Raw response (len=614): {
0|bot | “candidates”: [
0|bot | {
0|bot | “content”: {
0|bot | “role”: “model”
0|bot | },
0|bot | “finishReason”: “STOP”,
0|bot | “index”: 0,
0|bot | “groundingMetadata”: {}
0|bot | }
0|bot | ],
0|bot | “usageMetadata”: {
0|bot | “promptTokenCount”: 5366,
0|bot | “totalTokenCount”: 5540,
0|bot | “promptTokensDetails”: [
0|bot | {
0|bot | “modality”: “TEXT”,
0|bot | “tokenCount”: 1236
0|bot | },
0|bot | {
0|bot | “modality”: “VIDEO”,
0|bot | “tokenCount”: 3682
0|bot | },
0|bot | {
0|bot | “modality”: “AUDIO”,
0|bot | “tokenCount”: 448
0|bot | }
0|bot | ],
0|bot | “thoughtsTokenCount”: 174
0|bot | },
0|bot | “modelVersion”: “gemini-2.5-pro”,
0|bot | “responseId”: “81WfaOnvB8LRz7IP9bzs-QQ”
0|bot | }
0|bot |
0|bot | [DEBUG GEMINI:gemini-2.5-pro] Status: 200, OK: true
0|bot | [GROUP CHAT GEMINI] (gemini-2.5-pro) Lần thử 4/5…
0|bot | [DEBUG GEMINI:gemini-2.5-pro] Raw response (len=613): {
0|bot | “candidates”: [
0|bot | {
0|bot | “content”: {
0|bot | “role”: “model”
0|bot | },
0|bot | “finishReason”: “STOP”,
0|bot | “index”: 0,
0|bot | “groundingMetadata”: {}
0|bot | }
0|bot | ],
0|bot | “usageMetadata”: {
0|bot | “promptTokenCount”: 5366,
0|bot | “totalTokenCount”: 5380,
0|bot | “promptTokensDetails”: [
0|bot | {
0|bot | “modality”: “TEXT”,
0|bot | “tokenCount”: 1236
0|bot | },
0|bot | {
0|bot | “modality”: “VIDEO”,
0|bot | “tokenCount”: 3682
0|bot | },
0|bot | {
0|bot | “modality”: “AUDIO”,
0|bot | “tokenCount”: 448
0|bot | }
0|bot | ],
0|bot | “thoughtsTokenCount”: 14
0|bot | },
0|bot | “modelVersion”: “gemini-2.5-pro”,
0|bot | “responseId”: “AlafaOqkJYfVz7IPgoPNiAU”
0|bot | }
0|bot |
One thing I realized is if Gemini fails, don’t try to force it. Stop trying for few hours. When you try to force it few more times, you get partial responses which is worser than empty response.
1 Like
I’m facing the exact same issue here. Even with a robust retry mechanism that goes up to 10 attempts, we are still consistently failing to get a valid response. The API either returns a completely empty response or, what has become even more frequent for us, a 500 internal server error. What’s particularly frustrating is that everything was working perfectly stable for several months, but about a week ago, the reliability completely collapsed. Google, please fix this issue!!! The model is completely unusable in its current state.
1 Like
hope google was open about these updates. the issue is out there for months still there is no fix and workaround.
Hi @vjaykrsna ,
This issue has been escalated to our engineering team, and they are actively working on a resolution.
Thank you!
2 Likes
last whole week has been a disaster working with it. However over the weekend there were literally 0 issues, now Monday it literally failed on the first request…
This issue appears to be on certain accounts only, I’ve got several api keys. On one of my keys it’s failing almost 100% of the time, when I switch api keys and use the exact same prompt, it succeeds 100% of the time.
wow that seems to be a new news.
@Mrinal_Ghosh is this the case?
Very annyoing issue. It’s not about the API key, and the prompt.
Even with the same account and the same prompt, I sometimes get an empty response; a few seconds later, the exact same prompt returns a normal response.
my workaround is to re-send the request a few seconds after
I thought it was my fault. I combed through both the frontend and backend code and tried every method in the Gemini API docs, but nothing worked.
I’d prefer a clear failure signal instead.
1 Like