How does Gemini turn a minute of video or audio into billable tokens?

## Duration, not bytes The single most useful reframing here: for Gemini, media cost is a function of **time and dimensions**, never of file size. Re-encoding an hour of audio from 320 kbps to 64 kbps shrinks the upload by 80% and changes the token bill by exactly zero. A candidate who says "compress it to save tokens" has the wrong model of the pricing. ## The documented rates (Gemini 2.x families, mid-2026) - **Audio: ~32 tokens per second.** One minute is about 1,920 tokens; one hour about 115,000. - **Video: ~263 tokens per second** at default media resolution. One minute is about 15,800; ten minutes about 158,000. The figure reflects both the sampled frame and the accompanying audio. - **Images: 258 tokens** when both dimensions are at or below 384 px. Larger images are cropped into tiles, each resized to a fixed working size, and each tile costs 258 tokens — so a large screenshot costs a multiple of 258, not a smoothly scaling number. - **PDF pages** are treated like images: each page is documented at 258 tokens, with a cap of about 1,000 pages per document. A 40-page report is therefore roughly 10,000 tokens of page imagery before any output. These rates are stable enough to plan with, but they are model-family properties and have changed between generations — quote them with the family you are assuming, and check `count_tokens` rather than trusting memory in production. ## Why video is so much more expensive than audio Video is not "understood" as a stream. By default Gemini samples it at **one frame per second** and turns each sampled frame into image tokens, alongside the audio track at the audio rate. That is why the per-second video number is roughly eight times the audio one, and why a static talking-head recording costs exactly as much as fast-cut action footage of the same length — the sampler does not care that nothing moved. It is also why *duration*, not the resolution of the source, drives the bill: a 4K video and a 480p video of the same length cost the same, because frames are normalised before tokenisation. ## Context window pressure, not just money Token cost has two consequences and interviews usually probe the second one. At roughly 16,000 tokens per video minute, an hour of footage approaches a million tokens — it competes directly with everything else in the context window, and it slows the request down: prefill work scales with input tokens, so a long video means a long time to first token even before generation starts. This is why long-video workflows are usually segmented: analyse chunks, keep short structured summaries, and let the summaries rather than the raw media carry forward. ## Verifying instead of estimating `client.models.count_tokens(model=..., contents=[...])` accepts exactly the same multimodal contents as `generate_content`, including uploaded-file Parts, and returns `total_tokens`. Two practical uses: a pre-flight check so an over-budget request fails fast and locally rather than after a large upload, and a regression guard so a change in the documented rates shows up in your dashboards rather than in a surprise invoice. The response's `usage_metadata` closes the loop after the fact, reporting the prompt and candidate token counts the request was actually billed for. ## Estimating on the spot Useful round numbers to carry into an interview: **one minute of audio is about 2,000 tokens**, **one minute of video about 16,000 tokens**, **one image tile or PDF page 258 tokens**. From those, a 30-minute meeting recording is about 58,000 tokens of audio, but the same meeting as video is about 470,000 — an eightfold difference that usually settles the design question of whether you need the video at all. ## Where people go wrong Assuming file size drives the bill; assuming a lower-bitrate or lower-resolution source is cheaper; assuming silence or a static frame costs less; forgetting that video includes its audio track; and treating an image as a single flat 258 tokens regardless of size, when in fact anything much beyond 384 px on a side is billed as several tiles.