Gemini 4 Carbon supports video input and agentic video understanding. Static processing is the default. To have the model inspect the timeline, captions/transcript, and frames as needed, explicitly set processing: "agentic" on the video input.
Good for: Finding specific events in long videos, locating clips by topic, and cross-checking audio and visuals.
Not suitable for: Short clips where fast motion must not be missed or every frame must be inspected. Static mode's default 1 FPS may miss rapid changes, while agentic mode samples as needed.
Supported model version: Gemini 4 Carbon, model ID gemini-4-carbon.
Supported client, agent, or API: Gemini API Interactions API through the Google GenAI SDK.
Recommended reasoning level and parameters: Carbon defaults to thinking_level medium. For long videos, set processing: "agentic" on the video input. Use background=True or a streaming request for longer-running tasks. Video resolution can be configured independently.
The following Python snippet is adapted from the official agentic example. Replace the video path and query as needed. For larger videos, use the Files API and start the interaction after the file state becomes ACTIVE:
import time
from google import genai
client = genai.Client()
video_file = client.files.upload(file="path/to/video.mp4")
while not video_file.state or video_file.state.name != "ACTIVE":
if video_file.state and video_file.state.name == "FAILED":
raise RuntimeError("Video processing failed")
time.sleep(2)
video_file = client.files.get(name=video_file.name)
interaction = client.interactions.create(
model="gemini-4-carbon",
input=[
{
"type": "video",
"uri": video_file.uri,
"mime_type": video_file.mime_type,
"processing": "agentic",
},
{
"type": "text",
"text": "查找视频中讨论退款政策的片段,并按时间戳总结发言要点。",
},
],
background=True,
)
while interaction.status in ("queued", "in_progress"):
time.sleep(5)
interaction = client.interactions.get(interaction.id)
print(interaction.output_text)Safe timestamp extraction prompt example (English translation of an editorial example, not Google's original):
Based only on verifiable speech transcription or visuals in the video, list the segments related to "refund policy." For each, provide the start time (MM:SS), end time (MM:SS), evidence type (speech/visual), and a one-sentence summary. Do not infer information that does not appear. If the time or content cannot be confirmed, write "Unable to confirm" and do not invent a timestamp. When quoting directly, preserve the original wording and do not complete a cut-off sentence.Upload a large video or one you plan to reuse. A short, one-off video can also be passed directly using an officially supported input method.
Confirm that processing is complete, then explicitly set processing: "agentic" on the video item. If omitted, the default static mode is used.
Make queries specific to an event, topic, or time range. Use MM:SS format when locating a specific moment.
For long videos or multi-turn review, enable background execution or streaming. When finished, confirm the interaction status and read output_text.
Inspect interaction.steps. The presence of processing_call and processing_result indicates that the model dynamically inspected video content. Check extracted timestamps against the original video.
The Gemini 4 Carbon model card lists Agentic video understanding as supported and notes that static processing is the default.
The video guide lists Agentic as a mode supported by Carbon. Set processing: "agentic" on the video input to enable it. The model can inspect the timeline dynamically based on the prompt and read transcripts, frames, or audio as needed.
All Gemini models use static processing by default: one frame is extracted per second. The documentation says audio is processed as mono at 1 Kbps, with a timestamp added every second.
Google recommends static mode for short clips (about 5 minutes or less), tasks that need an overview of the entire video or frame-by-frame inspection, and latency-sensitive tasks. Agentic mode is suited to long videos (about 5–10 minutes or more) and searches for specific topics or moments. Google says it can reduce tokens by up to 88% for long videos. Short clips and broad questions may take more time or tokens because of navigation reasoning and tool round trips.
In Agentic mode, processing_call indicates a request to read a video segment or transcript, and processing_result is the corresponding result. Google recommends checking these steps to verify that dynamic navigation occurred.
Video clipping intervals and custom fps apply only to static processing. These parameters cannot control sampling in Agentic mode. In static mode, a prompt using MM:SS can locate content, but 1 FPS sampling may still miss fast motion.
Google documentation says 1M-context models can process videos up to 3 hours long at low resolution by default, or up to 1 hour at high resolution. Video resolution and processing mode are independent.
Dynamic navigation increases time to first token for long-video tasks. Use background execution or streaming for long tasks to avoid connection timeouts.
Google's token savings and quality improvements are capability claims in its documentation and are not guaranteed for every video or task.
A timestamp prompt can guide localization, but does not replace checking the original video. Clearly mark uncertainty when audio is missing, speech recognition is wrong, the video is edited, or evidence is insufficient.
No API call or video test was performed for this article. The code configuration is adapted from Google's official example.
Collected on 2026-11-16. The specified Google AI Studio video understanding documentation was opened through Tabbit and reviewed for its agentic mode, default sampling, limitations, and examples. The official Gemini 4 Carbon model card was also checked. Only the model ID in the code example was changed, and handling for a failed file state was added. No video was uploaded and no API request was made.
Gemini 4 Carbon