Articles
All posts, newest first. Each one carries a tag saying where its own figures came from — measured from records, or recalled from production without a reconciled ledger.
- Reference ad breakdown: seven fields per frame
A reference frame read into seven fixed fields, one frame per call. Batch the frames and the answers describe the set instead of the frame.
- AI image generation cost: what 9× buys
High quality cost 8.96× more per image and we could not reliably tell its renders from the cheap tier. The endpoint bills high by default.
- AI video revision cost: one rejected shot
A single rejected shot added ¥0.45 to a ¥1.80 job. The retry is cheap, but knowing when to retry is not, and that is where the real cost sits.
- Why AI video is random: five open slots
A video model fills in whatever you leave out. Give it five open slots and you get five coin flips. A start frame closes three of them.
- Why AI images get the text wrong
540 cells, two pipelines, the same reference images. The one difference we can attribute is a design choice: whether the model renders text at all.
- AI video pricing: the same model at 7.5×
One model, three channels, ¥0.11 to ¥0.83 per second. What changes is everything except the generated pixels: price, terms, and what you do yourself.
- Cost per AI ad is a denominator problem
200 shots cost ¥400 to generate. What that makes each finished ad cost depends on how many you choose to build, so no single figure is "the" cost.
- AI vs human ad creative: 26,013 creatives
Across 26,013 AI and 25,080 human creatives, cost per conversion matched. Then the edge vanished at scale, and the quality of the conversions lagged.
- What a 38.7-second AI ad actually costs
The itemised generation bill for a five-shot vertical ad, what a retry adds, and the two terms the number leaves out: labour, and reuse.