About ShotProbe

ShotProbe publishes hand-measured benchmarks for AI video and image generation tools. Everything here comes from production work, not from re-reading vendor pricing pages.

Who is behind this

I have worked in internet advertising for 15 years. For most of that time I have run creative production for paid social — the side that makes the ads, rather than the side that buys the inventory.

Today I run a 50-person creative production team. It used to be close to a hundred. That change is not because we produce less. It is because a growing share of what used to be built by hand is now generated, and the work that remains has shifted towards direction, selection and quality control. I have spent the last few years measuring that transition in production: what generation actually costs, and where it actually fails.

We ship creative to ByteDance, Tencent and Kuaishou — among the highest-volume paid-social inventory anywhere. At that rate, a difference of a few cents per second of finished video, or a retry rate that creeps from 20% to 30%, is not a curiosity. It shows up in the numbers that decide whether the team grows or shrinks.

Why this site exists

When we started using generative video seriously, almost every article about it repeated the same vendor-supplied facts: a duration range, a resolution, a headline price. None of them answered the questions that actually decided our workflow. Can a 40-second ad be generated in one call? What happens to the audio when you stitch segments together? Why did a shot that looked fine in preview come back with garbled on-screen text?

We answered those by hitting the walls ourselves, and we wrote the answers down. This site is that record, cleaned up.

How we test

• We run real projects. Costs come from the billing records of jobs we actually submitted, not from a pricing calculator. A figure like "¥3.42 for a 38.7 second ad" is an invoice line, not an estimate.
• We publish the failures. A rejected API call, a silently removed audio track, a duplicated text overlay — these are often more useful than the successes, and vendors rarely document them.
• We date our numbers and re-check them. Model prices, limits and availability change within months. Articles carrying figures state when they were measured, and we re-run the test before relying on them again.
• We label second-hand data. If a figure comes from someone else's published test rather than our own, the article says whose it is.

How we make money

Some links to tools are affiliate links, which means we may earn a commission if you sign up through them. This never changes what we recommend and never changes a number in a benchmark — the tools we pay for and use in production are the ones we write about, whether or not they pay anything. See theaffiliate disclosure for details.

Corrections

If a number here is wrong, we want to know — especially if you have a test that contradicts ours. Write to[email protected]. Corrections are published with a date and the original figure kept visible.