The Best AI for Making Videos in 2026
Our AI video generation guide is live, and this thread is about the number that decides your actual bill and appears on zero pricing pages: the retry tax. Every vendor advertises cents per second. Nobody advertises how many generations it takes to get one clip you would actually publish, and that ratio is the entire economics of this category. We tracked ours across a quarter of testing. Time to compare notes.
Full guide with the model rankings, the generation-first pipeline, the five-dimension self-test, and the responsibility layer is here: <https://whataidoineed.com/best/ai/for/making-videos>
**How the retry tax works:**
You prompt. The model returns something 80 percent right: the product label is gibberish, the hand has a bonus finger, the camera drifted somewhere you did not ask it to go. You re-prompt. And again. The published price says $0.40 per second; your finished 8-second hero shot cost five generations, so the real price was 5x the sticker, before counting your time. The vendor's showcase reel, meanwhile, is by definition the survivor of retries you never saw, which is why showcase quality and your first-attempt quality live on different planets.
**What our tracking showed (the shape, since exact ratios vary wildly by prompt type):**
The tax is not uniform across models, and it is not correlated with sticker price. Some cheaper models with lower per-generation costs needed so many attempts on hard prompts that they cost more per keeper than premium models. Some premium models nailed easy prompts in one but burned retries on exactly the same failure cases as everyone else.
The tax IS strongly correlated with prompt type. Abstract and atmospheric content: low tax, almost everything is usable because almost nothing is checkable. Text on objects, specific human actions, multi-element prompts, character consistency across shots: the tax climbs steeply, and this is precisely the content commercial work requires. The cruel summary: the retry tax is lowest on the content that matters least.
Prompt skill cuts the tax more than model choice does. The single biggest reduction we found came not from switching tools but from better generation plans: reference images as anchors, style blocks reused across prompts, and asking a text AI to write the video prompt before spending a single video credit. A dry run that costs nothing versus a wet run at $0.40 per second is the cheapest optimisation in the category.
**The buyer's takeaway:**
Evaluate every tool on cost per usable clip, never per-second rates, and measure it yourself during the free tier: one finished video, every generation logged, total spend divided by keepers. The guide's five-dimension test bench has this as the fifth dimension for a reason. It is the only number on the scorecard that is also your invoice.
**For the thread:**
Post your ratios: tool, prompt type, generations per keeper. Even rough numbers. A community retry-tax table sorted by prompt type would be genuinely unprecedented; no benchmark site publishes this because no vendor wants it published.
Share the tax-cutting tricks: the prompt patterns, reference-image workflows, or settings that measurably dropped your retries. Specifics over vibes.
And the confession corner: your most expensive single clip. How many generations, what finally fixed it, and was it worth it? The war stories are half the education in this category.