What should I use instead of the Google Video Intelligence API?
Use the Gemini API for most of what the Video Intelligence API did: labels, on-screen text, speech, explicit content and a description of each scene. Google recommends it as the replacement, and at about 100 tokens per second of video it costs a few cents per hour of footage on the cheaper Flash models. Keep a dedicated model for the jobs that need exact frames: shot boundaries, face detection and object tracks. Amazon Rekognition Video covers those as a managed API, and PySceneDetect or TransNet V2 cover shot boundaries for free.
Google announced the deprecation on September 14, 2026. The API keeps working until September 14, 2027, when it shuts down. Every pipeline that calls it has to move before then, and the move changes more than the endpoint, because a prompt-driven model returns different output from a fixed-vocabulary API.
What does the deprecation mean, exactly?
From the Video Intelligence API deprecations page: "Starting on September 14, 2026, Video Intelligence API is officially deprecated and will no longer be supported. You can continue using Video Intelligence API until September 14, 2027, when it will be shut down. We recommend migrating to Gemini family of models."
So existing calls keep returning results for another year, without support. Celebrity recognition already went away earlier, on September 16, 2025. Nothing in the notice says prices change during the wind-down, and the pricing page still lists the first 1,000 minutes of each feature free every month.
Which replacement fits each Video Intelligence feature?
| Video Intelligence feature | Price after 1,000 free min | Gemini API | Amazon Rekognition Video | Open source |
|---|---|---|---|---|
| Label detection | $0.10 / min | Yes, from a prompt with your label list | Label detection | CLIP or SigLIP zero-shot |
| Shot change detection | $0.05 / min | Approximate, to the nearest sampled frame | Segment detection (shots and technical cues) | PySceneDetect, TransNet V2 |
| Text detection (OCR) | $0.15 / min | Yes | Text detection | PaddleOCR on keyframes |
| Speech transcription | $0.048 / min | Yes, or Gemini 3.5 Transcribe | Amazon Transcribe (separate service) | Whisper |
| Explicit content detection | $0.10 / min | Yes, from a prompt | Content moderation | Open NSFW classifiers |
| Object tracking | $0.15 / min | No tracking output | Labels with bounding boxes | A detector plus a tracker, such as YOLO with ByteTrack |
| Logo detection | $0.15 / min | Yes, from a prompt naming the brands | Custom Labels on sampled frames (you train it) | A custom detector |
| Face detection | $0.10 / min | Not a face API | Face detection and face search | RetinaFace or InsightFace |
Google prices are from the Video Intelligence pricing page, checked November 2, 2026. Committed-use plans lower them to $0.08 to $0.12 a minute.
Anything that is a description of what happens fits the Gemini API well. Anything that has to land on an exact frame or follow a box across frames fits a dedicated model better.
How do I replace label detection with Gemini?
Send the video and ask for JSON with your own label list. The Gemini API samples one frame per second by default and reads the audio track too, so one call can return labels, on-screen text and a transcript summary with timestamps.
from google import genai
client = genai.Client()
video = client.files.upload(file="clip.mp4")
prompt = """Return JSON: a list of segments, each with start and end as MM:SS,
labels chosen ONLY from this list: [crowd, stadium, goal, interview, logo on screen, car, beach],
on_screen_text exactly as shown, and a one-sentence description."""
response = client.models.generate_content(
model="gemini-3.5-flash-lite",
contents=[video, prompt],
)
print(response.text)Three differences from the old API matter in production:
Your vocabulary replaces Google's. Video Intelligence returned entities from its own knowledge graph with confidence scores. Gemini returns the labels you ask for, which is usually better for search, but there is no confidence score unless you ask for one, and a model-reported score is not calibrated.
Timestamps follow the sampling rate. At the default one frame per second, a boundary lands to the nearest second. The
Output can vary between runs. Pin the model version, keep the prompt in version control and spot-check a sample after every model change.
Take 10,000 minutes of video a month with labels, shot changes and on-screen text, the most common Video Intelligence setup.
The Gemini rows assume about 100 tokens per second of video, the figure in Google's video understanding docs for the default resolution, and 500 output tokens per minute of video. High resolution is about 300 tokens per second, three times the input cost. Gemini 3.8 Flash prices double on January 1, 2027, per the Gemini API pricing page.
The Gemini rows buy you output text. None of the per-video APIs, Google's included, keep an index you can search across the library. If the old pipeline wrote annotations to BigQuery and people searched them there, that layer still has to exist after the move.
1. List what reads the annotations. Find every table, search index and dashboard that consumes Video Intelligence JSON. The consumers decide which fields you must keep, often only a few labels and the shot list. 2. Run both on a sample. Send 200 to 500 representative videos through the old API and the replacement. Compare label overlap, shot boundaries within one second and OCR text exactly. 3. Map the vocabulary. Write down the Google entity labels your consumers filter on and put those words into the new prompt, so downstream filters keep working. 4. Keep the old output. Store the last Video Intelligence results next to the new ones until the consumers have switched. After September 14, 2027 they cannot be regenerated. 5. Re-run the archive in batches. The Gemini Batch API costs less than interactive calls for back-catalog jobs that do not need an answer right away.
Mixpeek replaces the pipeline around the API as well as the API itself. It reads video from your own S3, GCS or Azure bucket, splits it into segments, extracts what is on screen, what is said and the text shown, and indexes all of it so one query returns timestamped moments across the library. That is the BigQuery-and-search layer most Video Intelligence users built by hand.
It fits when people search the footage. If you only need labels written into a table once, the Gemini API is cheaper. The footage intelligence template runs the search on a public sample library without an account, and pricing lists the per-minute rate.
On September 14, 2027. Google deprecated it on September 14, 2026, and it keeps working without support until the shutdown date.
The Gemini family of models. The deprecation notice says "We recommend migrating to Gemini family of models", and the Gemini API accepts video directly through the Files API, Cloud Storage or a YouTube URL.
For labels, text and descriptions, yes, by a wide margin. At about 100 tokens per second of video, Gemini 3.5 Flash-Lite input costs about $0.002 per minute, against $0.10 per minute for Video Intelligence label detection alone.
Roughly. It samples one frame per second by default, so a cut lands to the nearest second. For frame-exact cuts, use a shot detector such as PySceneDetect, TransNet V2 or Amazon Rekognition segment detection.
Most of it: labels, shot and technical-cue segments, text, content moderation and faces. Speech goes to Amazon Transcribe, a separate service, and logos need Rekognition Custom Labels, a model you train yourself and run on sampled frames.
They are yours and stay wherever you stored them. The API stops answering on September 14, 2027, so export anything you still need from it before then.
Google deprecated the Video Intelligence API on September 14, 2026, and it shuts down on September 14, 2027.
Google recommends the Gemini models, which cover labels, text, speech and explicit content from a prompt at about 100 tokens per second of video.
At 10,000 minutes a month, Gemini 3.5 Flash-Lite costs about $30 against $2,700 for Video Intelligence labels, shots and text.
Shot boundaries, faces and object tracks need a dedicated model: Rekognition Video, PySceneDetect, TransNet V2 or a detector with a tracker.
The API returned annotations and kept no index, so plan the search layer as part of the migration.
Best AI video analysis tools: the replacements compared on signals, price and data control
Best AI video tagging tools: label APIs, vision-language models and taxonomy tagging
What AI video analysis is and how it works: the five steps every tool runs
How video frame sampling works: what the sampling rate does to cost and recall
Twelve Labs vs Google Video Intelligence: a video-native API against the one being retired
Google Cloud, Video Intelligence API deprecations
Google Cloud, Video Intelligence API pricing
Google AI for Developers, Video understanding with the Gemini API
Google AI for Developers, Gemini API pricing
AWS, Amazon Rekognition Video and segment detection
PySceneDetect
Soucek and Lokoc, TransNet V2: An effective deep network architecture for fast shot transition detection
OpenAI Whisper
fps setting raises or lowers the rate, and the cost moves with it.What does each option cost at 10,000 minutes a month?
Take 10,000 minutes of video a month with labels, shot changes and on-screen text, the most common Video Intelligence setup.
| Option | How it is billed | Monthly cost |
|---|---|---|
| Google Video Intelligence | 9,000 billable minutes per feature x ($0.10 + $0.05 + $0.15) | $2,700 |
| Amazon Rekognition Video | $0.10 per minute for labels plus $0.05 for shot detection, from the AWS price list; text detection is billed on top | $1,500 before text |
| Gemini 3.5 Flash-Lite | 60M input tokens at $0.30 per 1M, plus about 5M output tokens at $2.50 per 1M | about $30 |
| Gemini 3.8 Flash | 60M input tokens at $0.75 per 1M, plus about 5M output tokens at $3.75 per 1M | about $64 |
| Mixpeek | $0.05 per minute of processed video, indexed and searchable | $500 |
The Gemini rows assume about 100 tokens per second of video, the figure in Google's video understanding docs for the default resolution, and 500 output tokens per minute of video. High resolution is about 300 tokens per second, three times the input cost. Gemini 3.8 Flash prices double on January 1, 2027, per the Gemini API pricing page.
The Gemini rows buy you output text. None of the per-video APIs, Google's included, keep an index you can search across the library. If the old pipeline wrote annotations to BigQuery and people searched them there, that layer still has to exist after the move.
How do I migrate without breaking what depends on the output?
1. List what reads the annotations. Find every table, search index and dashboard that consumes Video Intelligence JSON. The consumers decide which fields you must keep, often only a few labels and the shot list. 2. Run both on a sample. Send 200 to 500 representative videos through the old API and the replacement. Compare label overlap, shot boundaries within one second and OCR text exactly. 3. Map the vocabulary. Write down the Google entity labels your consumers filter on and put those words into the new prompt, so downstream filters keep working. 4. Keep the old output. Store the last Video Intelligence results next to the new ones until the consumers have switched. After September 14, 2027 they cannot be regenerated. 5. Re-run the archive in batches. The Gemini Batch API costs less than interactive calls for back-catalog jobs that do not need an answer right away.
Where does Mixpeek fit?
Mixpeek replaces the pipeline around the API as well as the API itself. It reads video from your own S3, GCS or Azure bucket, splits it into segments, extracts what is on screen, what is said and the text shown, and indexes all of it so one query returns timestamped moments across the library. That is the BigQuery-and-search layer most Video Intelligence users built by hand.
It fits when people search the footage. If you only need labels written into a table once, the Gemini API is cheaper. The footage intelligence template runs the search on a public sample library without an account, and pricing lists the per-minute rate.
Frequently Asked Questions
When does the Google Video Intelligence API shut down?
On September 14, 2027. Google deprecated it on September 14, 2026, and it keeps working without support until the shutdown date.
What does Google recommend instead of the Video Intelligence API?
The Gemini family of models. The deprecation notice says "We recommend migrating to Gemini family of models", and the Gemini API accepts video directly through the Files API, Cloud Storage or a YouTube URL.
Is the Gemini API cheaper than the Video Intelligence API?
For labels, text and descriptions, yes, by a wide margin. At about 100 tokens per second of video, Gemini 3.5 Flash-Lite input costs about $0.002 per minute, against $0.10 per minute for Video Intelligence label detection alone.
Can Gemini detect shot changes?
Roughly. It samples one frame per second by default, so a cut lands to the nearest second. For frame-exact cuts, use a shot detector such as PySceneDetect, TransNet V2 or Amazon Rekognition segment detection.
Does Amazon Rekognition do everything the Video Intelligence API did?
Most of it: labels, shot and technical-cue segments, text, content moderation and faces. Speech goes to Amazon Transcribe, a separate service, and logos need Rekognition Custom Labels, a model you train yourself and run on sampled frames.
What happens to my existing Video Intelligence annotations?
They are yours and stay wherever you stored them. The API stops answering on September 14, 2027, so export anything you still need from it before then.