Possible Bug in Gemini 2.5 Pro Behavior empyty response

Gemini API
, ,
1

Hi,

Has Gemini 2.5 Pro been nerfed recently, or is this a bug?

I’ve been working on an n8n automation project and noticed that recently, Gemini 2.5 Pro returns a lot of errors. These errors are unusual — there’s no warning or detailed message, just this output:

{
  "response": {
    "generations": [
      [
        {
          "text": "",
          "generationInfo": {
            "finishReason": "STOP",
            "index": 0
          }
        }
      ]
    ]
  },
  "tokenUsageEstimate": {
    "completionTokens": 0,
    "promptTokens": 15038,
    "totalTokens": 15038
  }
}

In n8n, this is not recognized as an error — it just returns with no text or output.

I have been using the exact same prompt before without issues, but this problem started occurring recently.

Extra Info: When using Gemini 2.5 Flash with the same prompt, this error does not occur.

How to Reproduce:

  1. Use Gemini 2.5 Pro via API or n8n with large prompt input (high token count).
  2. Run the workflow multiple times.
  3. this happen Occasionally before but increasingly often — the response will be empty with "completionTokens": 0 and "finishReason": "STOP".
2 Likes
3

Same. Lots of 500 errors starting 2-3 days ago, and beginning 20 hours ago I’m also getting empty responses that are not marked as errors, as well as truncated responses (well below the max token limit).
For a while it seemed that I can “avoid” this issue by using a much shorter prompt, but now I am getting empty/truncated responses even with the short prompt.

3 Likes
5

Same thing to me. I’m trying to use Tavily + Gemini 2.5 Flash and Pro to use Deep Research Automations and it happens a lot when Tavily returns PDF documents to process. When the workflow goes through ETL process it returns this:

image
image386×195 3.26 KB

So it doesn’t work using Flash or Pro model.
By the way, I changed the model from gemini-2.5-flash to
gemini-2.5-flash-preview-05-20 and it seemed to fix the issue. I’ll do more tests on this.
UPDATE
It failed the second time. It seems to be a bug. I hope it’s related to processing the same data two times and not related to processing great amounts of input tokens. If that’s the case I’m messed up.

{
    "response": {
        "generations": [
            [
                {
                    "text": "",
                    "generationInfo": {
                        "finishReason": "MAX_TOKENS",
                        "index": 0,
                        "citationMetadata": {
                            "citationSources": ""
                        }
                    }
                }
            ]
        ]
    },
    "tokenUsageEstimate": {
        "completionTokens": 0,
        "promptTokens": 226029,
        "totalTokens": 226029
    }
}
7

I am also facing the same issue. no response so i am changing it to 2.0 FLASH.

{

“response”: {

“generations”: [

[

{

“text”: “”,

“generationInfo”: {

“finishReason”: “STOP”,

“index”: 0

}

}

]

]

},

“tokenUsageEstimate”: {

“completionTokens”: 0,

“promptTokens”: 6091,

“totalTokens”: 6091

}

}