Batch API with gemini-3-flash-preview returns error 13 for video input (works with gemini-2.5-flash) #1890
Description
Environment details
- Programming language: Python 3.12
- OS: macOS
- Language runtime version: Python 3.12.4
- Package version: google-genai 1.56.0
Steps to reproduce
- Upload a video to the Files API
- Submit a batch job with
gemini-3-flash-previewreferencing the video - Wait for job completion
- Observe error 13 in the response
The same video works with gemini-2.5-flash in batch mode, and also works with gemini-3-flash-preview in sync mode.
Minimal reproduction script:
#!/usr/bin/env python3
"""
Minimal reproduction: Gemini 3 Flash batch API fails with video (error 13)
while Gemini 2.5 Flash works fine with the same video.
Requirements:
pip install google-genai yt-dlp
Usage:
export GEMINI_API_KEY=your_key_here
python repro.py
"""
import os
import subprocess
import tempfile
import time
from google import genai
from google.genai import types
def main():
api_key = os.environ.get('GEMINI_API_KEY')
if not api_key:
print("Set GEMINI_API_KEY environment variable")
return
client = genai.Client(api_key=api_key)
# Download a public YouTube video
print("Downloading test video from YouTube...")
tmp_dir = tempfile.mkdtemp()
yt_path = f"{tmp_dir}/test.mp4"
subprocess.run([
'yt-dlp', '-f', 'worst[ext=mp4]',
'-o', yt_path,
'https://www.youtube.com/watch?v=dQw4w9WgXcQ'
], capture_output=True, check=True)
# Encode to 5 seconds
print("Encoding to 5 seconds...")
encoded_path = f"{tmp_dir}/encoded.mp4"
subprocess.run([
'ffmpeg', '-y', '-i', yt_path, '-t', '5',
'-c:v', 'libx264', '-preset', 'fast', '-crf', '23',
'-r', '24', '-vf', 'scale=480:-2', '-an',
'-movflags', '+faststart', encoded_path
], capture_output=True, check=True)
print(f"Encoded video: {os.path.getsize(encoded_path) // 1024} KB")
# Upload video
print("Uploading to Files API...")
uploaded = client.files.upload(
file=encoded_path,
config=types.UploadFileConfig(display_name='test_video', mime_type='video/mp4')
)
video_uri = uploaded.uri
print(f"Uploaded: {video_uri}")
# Cleanup
os.unlink(yt_path)
os.unlink(encoded_path)
os.rmdir(tmp_dir)
# Simple request with system instruction and JSON schema
request = {
'contents': [{
'parts': [
{'file_data': {'file_uri': video_uri, 'mime_type': 'video/mp4'}},
{'text': 'Describe what you see briefly.'}
],
'role': 'user'
}],
'config': {
'system_instruction': {
'parts': [{'text': 'You are helpful. Be concise.'}]
},
'response_mime_type': 'application/json',
'response_schema': {
'type': 'object',
'properties': {
'description': {'type': 'string'},
},
'required': ['description']
},
},
}
# Test both models
models = ['gemini-2.5-flash', 'gemini-3-flash-preview']
jobs = []
print("\nSubmitting batch jobs...")
for model in models:
job = client.batches.create(
model=model,
src=[request],
config={'display_name': f'test_{model}'}
)
jobs.append((model, job.name))
print(f" {model}: {job.name}")
# Poll until complete
print("\nPolling for results...")
completed_states = {'JOB_STATE_SUCCEEDED', 'JOB_STATE_FAILED', 'JOB_STATE_CANCELLED'}
pending = set(range(len(jobs)))
while pending:
time.sleep(30)
for i in list(pending):
model, job_name = jobs[i]
job = client.batches.get(name=job_name)
state = job.state.name if hasattr(job.state, 'name') else str(job.state)
if state in completed_states:
pending.discard(i)
if hasattr(job.dest, 'inlined_responses') and job.dest.inlined_responses:
resp = list(job.dest.inlined_responses)[0]
if hasattr(resp, 'error') and resp.error:
print(f" {model}: ERROR {resp.error.code} - {resp.error.message}")
elif resp.response and resp.response.text:
print(f" {model}: SUCCESS")
else:
print(f" {model}: {state}")
if __name__ == '__main__':
main()Expected behavior:
Both models should successfully process the video and return a JSON response.
Actual behavior:
- gemini-3-flash-preview reports error 13.
- gemini-2.5-flash succeeds.
Downloading test video from YouTube...
Encoding to 5 seconds...
Encoded video: 165 KB
Uploading to Files API...
Uploaded: https://generativelanguage.googleapis.com/v1beta/files/[redacted]
Submitting batch jobs...
gemini-2.5-flash: batches/[redacted]
gemini-3-flash-preview: batches/[redacted]
Polling for results...
gemini-2.5-flash: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-2.5-flash: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-2.5-flash: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-2.5-flash: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-2.5-flash: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-2.5-flash: SUCCESS
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_PENDING
gemini-3-flash-preview: JOB_STATE_RUNNING
gemini-3-flash-preview: JOB_STATE_RUNNING
gemini-3-flash-preview: ERROR 13 - Internal error encountered.
Additional notes:
- Tested with multiple video encodings (baseline h264, various resolutions/framerates)
- Tested with and without system_instruction
- Tested with and without JSON schema
- Error occurs consistently (not intermittent)
- Job completes with
JOB_STATE_SUCCEEDEDbut response contains error 13 gemini-3-pro-previewhas the same issue