🎬 AI Video Token Calculator

A minute of video costs 5,880 tokens or 93,600, depending on where you send it

Video is the input where the token arithmetic stops being a rounding error. One minute of 1080p costs about 5,880 input tokens on Gemini and 93,600 if you extract the frames and send them to Claude as still images. That is the same minute of footage, sixteen times the bill, and the reason is not pricing — it is that only one of the two actually accepts a video file.

Only Gemini reads video natively

Google's API takes a video file directly. It samples at one frame per second by default and counts the audio track alongside the frames: 66 tokens per frame at the default media resolution, 258 at high, plus 32 tokens for every second of audio. Google states the combined rate as approximately 100 tokens per second of video, or approximately 300 at high. On a model with a one-million-token context window that allows up to three hours of video in a single request at the default setting, or one hour at high.

Gemini video tokens = seconds x (66 x fps + 32), or 258 per frame at high resolution

Written out that way because only the frame half scales with the sampling rate — the audio track costs 32 tokens per second whatever the frame rate. At the default one frame per second that formula gives 98 tokens per second at the default resolution and 290 at high, which is the roughly 100 and roughly 300 that Google publishes. The distinction matters as soon as you change the frame rate: halving it does not halve the cost, because the audio floor stays where it is.

Those two stated duration limits are really one limit wearing two hats. At 98 tokens a second, three hours comes to 1,058,400 tokens; at 290 a second, one hour comes to 1,044,000. Both sit just under a million-token context window, which is what is actually doing the limiting. The duration is the consequence, not the rule, and it is a neat confirmation that the per-frame and per-second components above are the right ones. If you need longer than that, you split the video into requests, and whether the model can reason across the join becomes your problem rather than the API's.

Google's own two pages disagree slightly. The video understanding guide gives 66 tokens per frame at low media resolution, 258 otherwise, plus 32 tokens per second of audio. The media resolution page gives 70 and 280 tokens per frame and 25 tokens per second of audio. The calculator above uses the first set, because those are the figures that add up to the roughly 100 and 300 tokens per second Google states as the totals. The gap between the two sets is three per cent at the default resolution and five per cent at high, so it will not change a decision, but treat any figure here as an estimate and verify against a real token count before committing to a budget.

Everywhere else, you are sending images

OpenAI and Anthropic accept still images only — PNG, JPEG, WebP and non-animated GIF. Anthropic's docs are explicit that an animated GIF contributes only its first frame. So “video input” on those platforms means you extract frames yourself with ffmpeg or similar and attach them as a sequence of images, and you pay the image rate for every one.

Anthropic publishes the exact arithmetic, which is unusual and useful. An image costs the ceiling of its width over 28, times the ceiling of its height over 28, in visual tokens — one token per 28-by-28 pixel patch. Before counting, the image is scaled down to the largest aspect-preserving size that fits both a maximum edge length and a visual token budget: 1,568 pixels and 1,568 tokens on the standard tier, 2,576 pixels and 4,784 tokens on the high-resolution tier available from Claude 4.7 onwards.

Claude image tokens = ceil(width / 28) x ceil(height / 28), after scaling

Read the first column against the third and the practical lesson is immediate: on the standard tier, 1080p, 1440p and 4K all cost exactly 1,560 tokens, because all three are scaled down to 1456x819 before counting. Uploading 4K frames to a standard-tier model buys you nothing at all. You pay the bandwidth, the encode time and the request size, and the model sees the same 1456x819 picture it would have seen from a 1080p source.

The high-resolution tier does reward more pixels, up to a point — 1080p rises to 2,691 tokens and 1440p saturates the 4,784 budget. Above 1440p it flattens again for the same reason. If detail genuinely matters, cropping the region you care about beats uploading a larger frame of the whole scene, because a crop spends its token budget on the part you want.

What the two routes cost on the same minute

The cheap Gemini row and the cheap Claude row land within a cent of each other, which is worth noticing before concluding that native video always wins on price: Haiku's token rate is low enough to absorb a sixteenfold token penalty. Where the gap bites is at the top of the range. A minute of video through GPT-6 Astra costs ninety-four cents of input against just over a cent on Gemini 4 Argon, and that is per minute, per request, before you have paid for a single output token.

For anything at volume the arithmetic compounds quickly. A thousand one-minute clips is 93.6 million input tokens on the frame route against 5.9 million natively. The pricing comparison will price that spread across every model at once.

Frame rate is the lever nobody pulls

Every figure above assumes one frame per second, which is Gemini's default and a sensible starting point. It is also frequently far more than the task needs. Identifying what a video is about, reading text on a slide, or checking whether a shelf is stocked does not need sixty frames for a minute of mostly-static footage. Drop to one frame every five seconds and the frame half of the bill drops by a factor of five.

The frame half, not the whole bill — and that is the part people get wrong. Dropping from one frame per second to one every five takes Gemini's rate from 98 tokens a second to 45, not to 20, because the 32 tokens of audio per second do not move. Below about one frame every two seconds the audio is the majority of what you are paying for, and cutting the frame rate further buys progressively less. If you are down there and still want the bill smaller, the thing to drop is the audio, not more frames.

Gemini lets you set a custom frame rate in the processing configuration, and on the frame route it is entirely your choice since you are doing the extraction. The other direction is worth knowing too: fast action, gesture recognition or anything where a single moment matters may need more than one frame per second, and the frame component scales linearly in that direction as well. Set the rate from what the task requires rather than accepting the default.

Real frame extractors deduplicate. If you pull frames with ffmpeg's scene-change filter rather than at a fixed interval, static footage produces a fraction of the frames a uniform sample would, and the token count falls with it. The calculator assumes uniform sampling because that is the worst case and the one you can predict; treat it as a ceiling.

Audio is cheap and often enough

Much of what people want from a video is in the soundtrack. Gemini counts audio at roughly 32 tokens per second on its own, against the 100 tokens per second that frames plus audio cost together — so stripping the video and sending audio alone is about a third of the price, and for a lecture, a meeting recording, a podcast or an interview it loses almost nothing. If you only need to know what was said, do not pay for the pixels.

A worked example

A ten-minute 1080p screen recording, sampled at one frame per second. On Gemini at the default resolution that is 58,800 input tokens, about 4.4 cents on Gemini 3.8 Flash. Extracted as 600 frames and sent to Claude Sonnet 5.5 it is 936,000 tokens, about $1.87, and it does not fit in a 200,000-token context window at all — so you would be splitting it into at least five requests, paying the prompt overhead each time, and losing the model's ability to compare the start against the end. The cost difference is just over forty times; the capability difference is larger than that.

Once you have this figure, the ai image cost is where it usually goes next. the ai token calculator is the related one people reach for afterwards.

Frequently Asked Questions (FAQs)