AI Studio video processing fails with HTTP 500 using a paid API key; text-only requests succeed

Google AI Studio
, , ,
1

I’m encountering repeatable video-processing failures in Google AI Studio. A text-only request succeeds with the same model and paid API key, but video imports and uploads fail.

Environment

  • Google AI Studio web interface

  • Google Chrome on macOS

  • Google AI Ultra account

  • Model: Gemini 3.8 Flash

  • Video requests: High thinking, Default media resolution, maximum output length 65,536

  • Paid API key linked through AI Studio’s prompt-box selector

  • Observed October 1, 2026, approximately 6:10–6:30 PM EDT

  • Browser automation was used to operate the AI Studio interface.

Steps to reproduce

  1. Open a new AI Studio prompt and select the paid API key.

  2. Select Gemini 3.8 Flash with High thinking.

  3. Attach a Google Meet MP4 through the Google Drive picker, or upload a local MP4.

  4. Enter a transcription prompt.

  5. Click Run.

Expected behavior

AI Studio processes the video and generates a transcript, or returns an actionable error explaining why it cannot process the input.

Actual behavior

The attachment displays “Token count failed.” The Run button remains enabled. Submitting the transcription request returns:

An internal error has occurred.

The browser console reports:

Error extracting video frames: RpcError:
Http response at 400 or 500 level, error: ,
http status code: 500

The video preview loads and reports the expected duration.

Tests performed

Test Result
Original recording attached through Google Drive Token counting failed; generation returned an internal error
A second recording attached through Google Drive in a fresh tab Token counting failed; console reported the same frame-extraction error
Byte-identical local copy of the first recording uploaded directly Upload completed; token counting and generation failed
First 30 seconds re-encoded as a separate MP4 and uploaded directly Token counting and generation failed
Text-only prompt, “Reply with exactly OK,” using the same key and model Returned “OK” successfully; this control used Medium thinking

The original recording is approximately 21 minutes 50 seconds and 222 MB, with H.264 video at 1280×720/24 fps and AAC audio. The short diagnostic file is exactly 30 seconds and approximately 7.5 MB, encoded as H.264/AAC with MP4 fast-start enabled.

Additional context

Before linking the paid API key, I was using AI Studio with a Google AI Ultra account and no API key selected. AI Studio displayed:

You’ve reached your quota for the day.

I then linked a paid API key to continue. Text-only requests succeeded, but video processing returned the internal errors described above. The quota message did not identify which quota had been exhausted, and I have not established whether the earlier quota issue and the video-processing failures are related.

Is this a known issue with AI Studio’s video preprocessing, or is there a project/API configuration requirement that could explain it? What additional diagnostics would help identify the cause?

2

Update: Switching from Gemini 3.8 Flash to Gemini 3.7 Flash did not resolve the issue.

Using High thinking and the same paid API key, I tested four different recordings through the Google Drive picker. All four returned “Token count failed” and HTTP 500 during attachment processing, before generation. No generation requests were submitted.