500 errors with many gemini models when audio input is used, since ~ 12:05 UTC

Gemini API
,
1

Suddenly gemini api request with audio content began to fail with 500 error:
Claude report:

model text image audio
gemini-3.5-flash-lite OK OK 500 INTERNAL
gemini-3.6-flash OK OK 500 INTERNAL
gemini-3.7-flash OK OK 500 INTERNAL
gemini-3.1-flash-lite OK OK OK
gemini-3.5-flash OK OK OK
gemini-3.1-pro-preview OK OK OK

I bisected the request feature by feature against both models, same 3-second audio, same prompt. On 3.5-flash-lite every audio variant 500s and every text variant passes:

  • streaming and non-streaming — both 500
  • thinkingLevel high / medium / low, thinkingBudget, no thinkingConfig at all — all 500
  • responseJsonSchema (nested transcript+analysis), flat schema, OpenAPI responseSchema, no schema at all — all 500
  • mp3 and wav, inline base64 and the Files API — all 500
  • text-only, with the identical config — 200
2

Same issue here…
I had a lot of issues on Friday as well.

What’s most striking, is that they never mention these issues on status pages or alike. Very frustrating.

3

It was back at about 16:00 UTC . 0 mention in any dashboard status, in the GCP console, nowhere.

4

Hi @OlivierWengo, we have attempted to reproduce this issue on our end but have been unsuccessful so far. To help us investigate this further, could you please clarify whether you are testing/using this via the Gemini API or Google AI Studio? Additionally, if possible, could you please share a sample prompt or a minimal reproducible code snippet containing the audio input structure you are using?

5

@Paypal_Sharma2 The issue lasted for about 4 hours on August, 17th . Many gemini-3.x-flash and gemini-3.x-flash-lite made a “500 INTERNAL ERROR” when an audio input was provided. Claude generated many configurations with and without structured output, image input etc, the result was that it was the audio modality which caused the outage. Since we use these models for transcriptions and analysis we were impacted hard and the only fallback was gemini-3.1-flash-lite ( the pro-preview is too expensive and too slow ).
There was a period with a high rate of errors but no more 100% , and then everything came back to normal.