One YouTube link, read nine ways: from the key shots to the transcript, each priced

A transcript is one thing a video holds. It also has a look, a palette, a sequence of shots, people with names and faces, text on screen and a soundtrack. An agent researching a competitor’s launch video, a creator’s style or a product demo needs some of those and rarely all of them.

youtube describe is a program in the shell your agent gets from hedwigai. Give it a public YouTube link and a model watches the video and answers in JSON, with only the parts you ask for. We ran each of its nine parts on our own two-minute explainer, then priced the look of a fifteen-minute speech on four models, on 3 November 2026.

Most of every row is the video itself. Before it writes a word, the model reads the whole thing as input tokens, about 11,000 for these two minutes, so each of the eight aspects on the lighter model costs about seven tenths of a cent whether it writes 72 tokens or 1,752. Asked one by one, the nine cost $0.0844; asked together, $0.0239. Ask for everything you will need in one call.

--aspects visuals returns the style in two sentences, the main colours by name, and one entry per shot: when it starts, the kind of shot, what is in frame, where things sit and the light.

That is a storyboard of the explainer, written from the frames, for under a cent. On the Stanford commencement address, fifteen minutes of a man at a podium, the same call described “a medium shot of the speaker at a podium and occasional cutaways to the audience”, lit “bright and natural, consistent with an outdoor daytime event”, in red, white, black and brown, across 15 key shots from the Stanford logo to the graduates clapping.

Gemini’s 2.x models read a video at about 297 tokens a second, a frame each second plus the audio. The Gemini 3 models read the same video at about 88. On a fifteen-minute speech that is 267,185 tokens against 79,795 before any answer is written, which is why the model matters more than the question on a long video.

gemini-3.1-flash-lite gave a usable look for a third of what 2.5-flash charged, and 2.5-flash ignored the cap of 40 shots it was given, so the program now cuts the list at 40 itself. Its transcript was another matter: it voiced the on-screen title as the narrator’s first line and put a sentence at 0:13 that is not spoken there. So the program chooses for you when you do not name a model: gemini-2.5-flash for an answer with a transcript, gemini-3.1-flash-lite for any other. Name one with --model to override it.

The story comes back as beats with their times, which reads as a chapter list. Characters are named only when the video names them, so our narrator is “a man with a beard and sunglasses”. On-screen text is the one light aspect that writes a lot: our explainer scrolls an article, and the model copied it down, timed, in 1,752 tokens.

--transcript prints only the transcript, a line per sentence or two, and asks for nothing else, so nothing else is paid for:

That run spelled every name the way the screen does. Not every run did. In two of the four transcripts we took of this video, the model wrote names the way they sound: “Cloud code” for Claude Code, “deep seed V3s”, “a birthly paper” for a Berkeley one, and “the head of block” for our own blog.

So the transcript is now told to spell names, products and technical terms as the screen shows them or as they are usually written, never as they sound. Run again on the same model, the same call wrote Claude Code, DeepSeek-V3, “a Berkeley paper”, Qwen and Mac, for the same $0.018. The one name it still heard was ours, which the video says but never shows: “the Headwigga blog”. A name that is only spoken is still worth checking.

Sound was the weakest aspect. A music bed runs under the narration the whole way through, and no gap in the audio falls below -30 dB, yet three of the four answers we checked said there was no music. Only the first run heard “calm, ambient background music”. Treat sound as a hint.

Watching takes a minute or two, and not in proportion to the length: the two-minute explainer took a little under two, the fifteen-minute speech about a minute and a half. Many MCP clients drop a tool call after about 100 seconds, so the program does not hold the call open. It starts the watching as a job, waits up to 80 seconds, and if the answer is not ready it says how to pick it up: