AI That Watches Video Can Be 85% Cheaper, Study Finds

If you've ever paid for AI tools, you've paid in "tokens" — the units these systems charge you for, a bit like paying by the word. Vision-language models, AI that can watch video and read text at the same time, burn through tokens fast. Watching a typical 60-second social video one frame per second can cost millions of tokens. That's like paying a translator by the second for a two-hour movie.

So researchers at arXiv tested popular money-saving tricks: slow the video down to fewer frames, squeeze frames into a single image grid, or skip video entirely and just use the text. They ran these on two real research tasks — figuring out whether videos were positive or negative, and what they were about — and measured three things: how accurate the AI was, whether the conclusions actually held up, and what each video cost.

The results were surprising in two ways. First, accuracy and truth can disagree. The setting with the best-looking accuracy scores still produced wrong overall conclusions, which matters if you're using AI to make business or research decisions. Second, adding video to text isn't automatically better — plain text alone performed strongly, meaning extra video frames can add cost without adding insight.

The big win: a single 2×8 image grid, assembled by simple scene-change detection, matched full-video understanding almost exactly while costing roughly 15% as much. In plain terms, showing the AI a smartly chosen collage of key moments works nearly as well as showing it everything. The takeaway for anyone paying for AI: more expensive isn't automatically more accurate, and asking the right questions before you spend is where the real savings live.