Most writing about AI video is a feature announcement wearing a blog post. It opens with a paragraph about how everything is changing, lands on a product screenshot, and never once tells you what anything cost or how long it took.
We are going to try to do the opposite here.
The rule we are holding ourselves to
Every speed claim comes with the artefact it was measured from: the video, the timer, the machine it ran on. Every price is dated, because rate cards move and a number without a date is a rumour. Every comparison names the version of the tool we tested, because “we tried it and it was worse” is not a finding.
That rule has a cost, and it is worth being upfront about it: it means we will publish less often than a blog that does not hold itself to it, and it means some posts will end with “we do not know yet” rather than a conclusion.
What we will write about
There are a handful of subjects where we have numbers other people do not, mostly because we are the ones running the pipeline every day. Those are the only things worth our publishing.
What video actually costs. Agency rate cards, freelance quotes and credit-tool pricing, priced against each other at a real cadence rather than per video. One video is a bad unit of comparison. Twelve videos a year is a budget.
How the agent pipeline works. What an AI agent does well when it builds a video, where it fails, and what a project file has to look like for the agent to be able to change it later. That last one is the part nobody writes about and it is the part that decides whether the tool is usable in month three.
Measurements we ran ourselves. Render times, transcription speed on WebGPU against the WebAssembly fallback, how long a real sixty-second demo took end to end. Timestamped, with the hardware named.
Where competing tools beat us. Stated first, in the post, not buried in a final paragraph. A comparison that never concedes anything is an advertisement, and it reads like one to anybody evaluating seriously.
What the agent gets wrong. The failure modes worth knowing before you rely on one: where a brief is too vague to build from, which corrections take one turn and which take five, and the kinds of video an agent still should not be handed.
What we will not do
We will not publish a benchmark we cannot reproduce. We will not compare our current version against a competitor’s release from last year. We will not run a post whose only argument is that AI is changing everything.
If we cannot show you the artefact, we do not get to make the claim.
If you find an error
Some of what we publish will turn out to be wrong. When it does, the correction goes at the top of the post with the date it was made, and the original claim stays visible underneath it. Quietly editing a number after the fact is worse than having got it wrong in the first place.