Gemini 2.5 Pro Outputs Gibberish (via Vertex AI)

[wjkoh]

Is there anyone who’s experiencing an issue with Gemini 2.5 Pro now? Suddenly, Gemini 2.5 Pro via Vertex AI started returning empty or gibberish values for JSON output as follows:

2025/08/11 17:19:02 INFO transcribeBytes respText="[\n  {\n    \"time\": \"00:01\"\n  },\n  {\n    \"time\": \"00:01\"\n  },\n  {\n    \"time\": \"00:06\"\n  },\n  {\n    \"time\": \"00:10\"\n  },\n  {\n    \"time\": \"00:14\"\n  },\n  {\n    \"time\": \"00:19\"\n  },\n  {\n    \"time\": \"00:22\"\n  },\n  {\n    \"time\": \"00:25\"\n  },\n  {\n    \"time\": \"00:34\"\n  },\n  {\n    \"time\": \"00:38\"\n  },\n  {\n    \"time\": \"00:43\"\n  },\n  {\n    \"time\": \"00:48\"\n  },\n  {\n    \"time\": \"00:54\"\n  },\n  {\n    \"time\": \"01:05\"\n  },\n  {\n    \"time\": \"01:11\"\n  },\n  {\n    \"time\": \"01:19\"\n  },\n  {\n    \"time\": \"01:32\"\n  },\n  {\n    \"time\": \"01:37\"\n  },\n  {\n    \"time\": \"01:41\"\n  },\n  {\n    \"time\": \"01:46\"\n  },\n  {\n    \"time\": \"01:48\"\n  },\n  {\n    \"time\": \"01:52\"\n  },\n  {\n    \"time\": \"01:56\"\n  },\n  {\n    \"time\": \"02:00\"\n  },\n  {\n    \"time\": \"02:02\"\n  },\n  {\n    \"time\": \"02:06\"\n  },\n  {\n    \"time\": \"02:14\"\n  },\n  {\n    \"time\": \"02:20\"\n  },\n  {\n    \"time\": \"02:29\"\n  },\n  {\n    \"time\": \"02:32\"\n  },\n  {\n    \"time\": \"02:36\"\n  },\n  {\n    \"time\": \"02:40\"\n  },\n  {\n    \"time\": \"02:53\"\n  },\n  {\n    \"time\": \"02:57\"\n  },\n  {\n    \"time\": \"03:02\"\n  },\n  {\n    \"time\": \"03:07\"\n  },\n  {\n    \"time\": \"03:12\"\n  },\n  {\n    \"time\": \"03:20\"\n  },\n  {\n    \"time\": \"03:28\"\n  },\n  {\n    \"time\": \"03:30\"\n  },\n  {\n    \"time\": \"03:34\"\n  },\n  {\n    \"time\": \"03:38\"\n  },\n  {\n    \"time\": \"03:42\"\n  },\n  {\n    \"time\": \"03:53\"\n  },\n  {\n    \"time\": \"04:00\"\n  },\n  {\n    \"time\": \"04:05\"\n  },\n  {\n    \"time\": \"04:08\"\n  },\n  {\n    \"time\": \"04:12\"\n  },\n  {\n    \"time\": \"04:18\"\n  },\n  {\n    \"time\": \"04:23\"\n  },\n  {\n    \"time\": \"04:27\"\n  },\n  {\n    \"time\": \"04:37\"\n  },\n  {\n    \"time\": \"04:42\"\n  },\n  {\n    \"time\": \"04:46\"\n  },\n  {\n    \"time\": \"04:54\"\n  },\n  {\n    \"time\": \"04:56\"\n  },\n  {\n    \"time\": \"04:57\"\n  },\n  {\n    \"time\": \"05:02\"\n  },\n  {\n    \"time\": \"05:06\"\n  },\n  {\n    \"time\": \"05:08\"\n  },\n  {\n    \"time\": \"05:10\"\n  },\n  {\n    \"time\": \"05:14\"\n  },\n  {\n    \"time\": \"05:17\"\n  },\n  {\n    \"time\": \"05:25\"\n  },\n  {\n    \"time\": \"05:29\"\n  },\n  {\n    \"time\": \"05:32\"\n  },\n  {\n    \"time\": \"05:38\"\n  },\n  {\n    \"time\": \"05:40\"\n  },\n  {\n    \"time\": \"05:47\"\n  },\n  {\n    \"time\": \"05:52\"\n  },\n  {\n    \"time\": \"05:54\"\n  },\n  {\n    \"time\": \"06:02\"\n  },\n  {\n    \"time\": \"06:06\"\n  },\n  {\n    \"time\": \"06:09\"\n  },\n  {\n    \"time\": \"06:13\"\n  },\n  {\n    \"time\": \"06:19\"\n  },\n  {\n    \"time\": \"06:24\"\n  },\n  {\n    \"time\": \"06:28\"\n  },\n  {\n    \"time\": \"06:31\"\n  },\n  {\n    \"time\": \"06:36\"\n  },\n  {\n    \"time\": \"06:43\"\n  },\n  {\n    \"time\": \"06:47\"\n  },\n  {\n    \"time\": \"06:56\"\n  },\n  {\n    \"time\": \"07:02\"\n  },\n  {\n    \"time\": \"07:05\"\n  },\n  {\n    \"time\": \"07:08\"\n  },\n  {\n    \"time\": \"07:13\"\n  },\n  {\n    \"time\": \"07:16\"\n  },\n  {\n    \"time\": \"07:20\"\n  },\n  {\n    \"time\": \"07:25\"\n  },\n  {\n    \"time\": \"07:30\"\n  },\n  {\n    \"time\":

[Mrinal_Ghosh]

Hi @wjkoh ,

Welcome to the Forum!!
From my end, the functionality seems to be working correctly based on my own test inputs. To ensure we are aligned, would you mind sharing the prompt you are using?

[wjkoh]

thanks for checking! my prompt includes a single, 14MB mp3 file (~30min long) and the prompt is “Transcribe the provided audio.” I’m retrying now but getting the following error

transcribe_test.go:66: doRequest: error sending request: Post "https://aiplatform.googleapis.com//v1beta1/projects/projectnamehere/locations/global/publishers/google/models/gemini-2.5-pro:generateContent": http2: client connection lost

[wjkoh]

I got max_tokens after all those empty jsons

[wjkoh]

another failed attempt with max_tokens

[wjkoh]

The issue still persists with GitHub - googleapis/go-genai: Google Gen AI Go SDK provides an interface for developers to integrate Google's generative models into their Go applications. v1.19.0.

[wjkoh]

[It turns out that this details are unrelated.]

[wjkoh]

After a few tries, I was able to fix the issue by setting the temperature to zero and adding property ordering. I’m not sure whether this change fixed it or the model is fixed during the meantime. The generation time is too long for 30 min audio, so I don’t have time to test if the old setting still cause the issue. Anyways, I’m moving on and thanks for the help!