How I cut the AI cost of a lesson by about 70%

A lesson in Unwatch is made by an AI model that watches a YouTube video and writes it up: the spoken content, plus the code and text shown on screen. When I started, that cost about ₹11.7 per hour of video. Today it costs about ₹3 to ₹3.7, with the same on-screen code and text. This is how I got there, including the ideas that didn't work.

All prices below include 18% GST and use ₹88 to the dollar. They come from Google's price pages as I read them on 15 October 2026.

Why the cost mattered

I wanted a paid plan around ₹199 a month that includes many hours of video. At ₹11.7 an hour, a user who used every hour would cost me more than they paid. Google also lists a price rise for the model I was using: input tokens go from $0.75 to $1.50 per million on 1 January 2027, which would have taken my cost to about ₹23 an hour.

Where the money goes

I measured what a YouTube video costs Gemini to read. At a low frame rate (0.1 frames per second) it is about 31.6 tokens a second, around 114,000 tokens per video-hour. Frames are cheap: about 66 tokens each. Audio is about 25 tokens a second, which is 79% of the input. You pay for the audio whenever the model looks at a clip, however few frames you send.

That one number rules out many cheap-sounding ideas.

What I tried

I benchmarked more than ten setups on the same clips. The main three: a 15-minute code-heavy API tutorial, a 2-minute fast-cut React video, and a 10-minute whiteboard lecture. I counted "screen facts": things only visible on screen, such as an exact URL.

On the other two clips, the chosen setup cost ₹4.69 per hour against ₹19.90 (React), and ₹3.08 against ₹12.71 (whiteboard). That's a cut of 72 to 76% on all three, and the on-screen facts matched the expensive model. Counting everything (key takeaways, search embeddings, short videos) my planning figure is about ₹3 per video-hour.

What made the difference

1. A cheaper model with a better prompt. Flash-Lite alone wrote far fewer code blocks than Flash: 3 against 12 on the code clip. A prompt with a rule about copying code and board text exactly, and a short example of the format, brought most of it back at almost no extra cost. For small models, the prompt mattered more than frame rate or thinking settings.

2. Shorter clips. Splitting the video into 5-minute clips, and passing the headings already written to the next clip, took the code blocks from 3 to 14 and the screen facts to 6 of 6. It cost about a third more than 15-minute clips, but the lessons are longer and complete.

3. The Flex tier. Flex costs half of the standard price and runs as a normal request, with no waiting like a batch job. It is best effort: it can return a 503 when busy, so the app retries and falls back to the standard tier. The clips I measured took 5 to 8 seconds each.

4. Output is the cheap part. On Flash-Lite, about 80% of the cost is input, mostly audio. Asking for longer, fuller lessons costs little, so I stopped trying to make them shorter.

What didn't work

• Transcripts instead of video. Captions are cheaper, but the lesson had 0 of 6 screen facts and wrote the spoken URL instead of the one on screen. I wrote about that separately: why summarisers miss on-screen code. Getting captions also means scraping YouTube, which its terms don't allow; Gemini reading the video URL is a documented feature.
• Looking at the screen only where needed. This sounds cheaper, but the audio comes with every window, and a second of high-detail video costs about three times as much as a second at low detail. Looking at 186 seconds in detail cost about as much as watching all 10 minutes lightly. The small models I used to pick the moments also missed many of them.
• Medium resolution, a word budget and low thinking. Medium resolution used exactly the same tokens. A word budget saved 8% but dropped facts. Low thinking made the lessons shorter and worse.
• Telling the checker what to expect. When I told the model the spoken URL as a hint, it "read" that wrong URL off the screen. Hints must describe the kind of thing, not the answer.

What I got wrong along the way

I assumed my API key was on the free tier, so for a couple of days I wrote "₹0" next to my test runs. I later found billing was on, and about ₹29 had been charged. It was small, but I now run all tests on a separate free key and keep the paid one for production. Every AI call is logged with its tokens and cost, so I can check the real numbers.

What I haven't verified

• Each setup ran once on three clips. Model output varies, so small differences mean little. The 70% cut is large enough that I'm confident in it.
• The Flex price is Google's published price. I haven't yet confirmed on an invoice that it bills at half.
• My own lessons from the production log cost about ₹0.50 per 5-minute clip at the standard rate (₹6 an hour). The half-price Flex tier should bring that to about ₹3.