—input tokens / video
—input tokens / minute
—monthly cost
Frame rate is the dial — sweeping it
Image tokens scale directly with frames sampled. For slow footage — lectures, screen recordings, security feeds — a lower frame rate rarely loses anything and cuts the bill in proportion. Audio and prompt are held at your settings.
| Sampling | Input tokens / video | Cost / video | Monthly |
|---|
The modality nobody budgets for
Text is cheap, images are pricier, and video is in a category of its own — because a model turns a video into a wall of image tokens, one sampled frame at a time, and stacks the audio track on top. The intuition trap is thinking of a video as "one input" like a document; in reality a ten-minute clip at one frame per second is 600 frames, roughly 155,000 image tokens plus 19,000 audio tokens, and a full hour crosses a million input tokens before the model produces a single word of output. That is why a video-understanding feature that looked trivial in a demo can dominate an entire AI bill once it runs across a library. The levers are direct: sample fewer frames per second, drop the audio when you only need the visuals, transcribe speech with a cheap speech-to-text model and reason over the text instead of paying audio-token rates, and clip to the segment that matters. Size the transcript alternative on the speech-to-text cost calculator, the still-image side on the vision image input calculator, and the generation direction on the video generation cost calculator.
How this calculator works
It counts the sampled frames (length in seconds × frames per second), multiplies by the tokens per frame to get image tokens, adds audio tokens (seconds × audio tokens per second) and your prompt to get total input tokens, then prices input and output at your per-million rates. Cost per video times your monthly volume gives the monthly bill, and the table re-runs the whole thing across a range of frame rates so you can see the trade-off.
Frequently asked questions
How is video billed when you send it to an LLM?
A multimodal model does not watch video the way a person does — it samples frames, usually about one per second, and treats each frame as an image made of tokens, then adds the audio track as its own stream of tokens. So the input cost is the number of sampled frames times the tokens per frame, plus audio tokens, plus your text prompt. Gemini bills roughly 258 tokens per second of video at default resolution and around 32 tokens per second of audio; with GPT-4o you extract frames yourself and each tiled frame runs from about 85 to over 1,000 tokens depending on detail. Either way, video is by far the most token-hungry input modality, which this calculator makes concrete.
Why is a one-hour video so expensive to analyse?
Because the tokens scale with the length of the footage. At one frame per second and 258 tokens per frame, an hour of video is 3,600 frames and about 929,000 image tokens, plus roughly 115,000 audio tokens — well over a million input tokens for a single video, before the model writes a word of answer. At a few dollars per million input tokens that is a meaningful bill per video, and it multiplies fast across a library. The calculator shows the tokens per video, the cost per video, and the monthly total so a one-off experiment does not turn into a surprise at scale.
How do I make video understanding cheaper?
The biggest lever is frame rate: sampling at one frame every two or five seconds instead of one per second cuts image tokens proportionally, and for slow-moving footage — a lecture, a security feed, a screen recording — it rarely loses anything that matters. Dropping the audio track when you only need the visuals, or transcribing audio with a cheap speech-to-text model and feeding the transcript as text instead of paying audio-token rates, is another large saving. Lowering frame resolution reduces tokens per frame, and clipping to the relevant segment rather than sending the whole file is the simplest win of all. The frame-rate table on this page shows exactly how much each setting moves the bill.
Is it cheaper to transcribe the audio than to send the whole video?
Very often, yes, if your question is about what is said rather than what is shown. Speech-to-text models charge per minute of audio at a fraction of a cent, and the resulting transcript is a few thousand text tokens for an hour of speech — orders of magnitude cheaper than paying image-token rates on a frame every second. Send the full video only when the visual content genuinely matters; for talking-head footage, meetings, and podcasts, transcribe first and reason over the text. This calculator lets you set audio tokens to zero to model the transcript route and compare it against sending the frames.