How many tokens is an hour of video?
In Gemini, one hour of video is about 352,800 tokens at low media resolution and 1,044,000 tokens at the default resolution. Gemini samples one frame per second and counts 66 tokens a frame at low resolution or 258 at default, plus 32 tokens a second for the audio track. On Gemini 2.5 Flash that hour costs about $0.19 of input at low resolution and $0.39 at default.
Claude takes video as still frames. A 1280x720 frame is 1,196 tokens, so one frame a second for an hour is 4,305,600 tokens, more than any single request holds, and the audio needs a separate transcript.
How does each model count video tokens?
We read each provider's documentation and pricing page on 2026-10-24.
| Gemini | Claude | |
|---|---|---|
| Video file input | Yes, sampled at 1 frame a second by default | No, send frames as images |
| Tokens per frame | 258 (66 at media_resolution low) | ceil(width / 28) x ceil(height / 28) |
| Audio | 32 tokens a second, read from the file | Not accepted; transcribe it first |
| Change the frame rate | fps in the processing config, such as 0.5 | Choose how many frames you extract |
| Longest video in one request | About 3 hours at low resolution, 1 hour at default (1M context) | Up to 600 images a request |
What does one hour of video cost to send?
Input cost only, at standard list prices on 2026-10-24. For Gemini we priced the audio track at the model's audio rate where one is listed.
| Model and setting | Tokens for one hour | Input cost |
|---|---|---|
| Gemini 2.5 Flash-Lite, low resolution | 352,800 | $0.06 |
| Gemini 2.5 Flash, low resolution | 352,800 | $0.19 |
| Gemini 2.5 Flash, default resolution | 1,044,000 | $0.39 |
| Gemini 3.1 Flash-Lite, low resolution | 352,800 | $0.12 |
| Gemini 2.5 Pro, low resolution | 352,800 | $0.88 |
| Claude Sonnet 5.5, 720p at 1 frame a second | 4,305,600 | $8.61 |
| Claude Sonnet 5.5, 360p every 5 seconds | 215,280 | $0.43 |
| Claude Haiku 5.5, 360p every 5 seconds | 215,280 | $0.11 |
Gemini 2.5 Pro bills prompts over 200,000 tokens at $2.50 per million, which an hour of video always crosses. Claude Haiku 5.5 is $0.10 per million for prompts up to 100,000 tokens and $0.50 above, so batching frames into smaller requests lowers its rate. The script that produces every number here is short enough to rerun with your own frame size and rate.
Why does asking questions about a long video get expensive?
A chat or an agent that keeps the video in the prompt sends every token again with each question. Asking 100 questions about the same hour on Gemini 2.5 Flash at low resolution is 100 x $0.19, about $18.65 of input.
Two ways to stop paying for the whole hour each time:
1. Cache the video. Gemini's context caching bills cached video tokens at $0.03 per million on 2.5 Flash, plus $1.00 per million tokens per hour of storage. The same 100 questions come to about $1.86 of cached reads, plus $0.35 for each hour the cache stays alive. 2. Index the video once and retrieve. Split it into segments, store a caption, a transcript and an embedding for each, and send only the segments that match the question. Ten segments of about 250 tokens each is 2,500 tokens a question, so 100 questions cost about $0.08 of input after a one-time $0.19 pass.
The index also answers questions across many videos, which no context window holds. A library of 1,000 hours is about 353 million tokens at Gemini's low resolution.
How do I cut the token count for a video?
Doing it in Mixpeek
Mixpeek reads your files where they already live, in S3, GCS or any bucket, and turns what's in them into data your software can search, classify and moderate. Its multimodal extractor splits each video into segments (every 10 seconds by default, or at scene changes or pauses), then stores a Whisper transcript, a Gemini description, on-screen text and an embedding for each segment. A retriever finds the segments that match a question, and your model reads those segments with their timestamps.
Who it fits: teams asking many questions over many videos, such as media archives, ad libraries and meeting recordings. Who it does not: if you ask one question about one short clip, sending the clip straight to Gemini is simpler.
Frequently Asked Questions
How many tokens is one minute of video in Gemini?
About 5,880 tokens at low media resolution (60 frames at 66 tokens plus 1,920 audio tokens) and 17,400 at the default resolution (60 frames at 258 tokens plus 1,920 audio tokens), at Gemini's default rate of one frame a second.
Can Claude watch a video?
The Claude API takes images and text and has no video or audio input. To ask Claude about a video, extract frames, send them as images (up to 600 a request), and add a transcript of the audio as text. Each frame costs ceil(width / 28) x ceil(height / 28) tokens.
How long a video fits in Gemini's context window?
Google's documentation says a model with a 1M token context window takes about 3 hours of video at low media resolution or about 1 hour at the default resolution.
Is it cheaper to index video or send it straight to an LLM?
For one question about one video, sending it straight to the model is cheaper and simpler. Once you ask more than a handful of questions, or ask across a library, indexing wins: each question reads a few thousand tokens of retrieved segments, where the whole video is hundreds of thousands.