How Long Does a Generation Actually Take? We Measured a Month of Ours

Every generator on the internet says it is fast. Very few say how many seconds, which is convenient, because seconds can be checked and adjectives cannot. So here are ours.

The numbers below come from our own generation rows for the thirty days ending 23 September 2026 - the same rows that decide what you were charged, not a benchmark run on a quiet afternoon. They are a description of one month on one service, not a promise about your next click.

Images: about a minute, with a tail

Over that window there were 1,533 image attempts. 1,525 finished successfully and 3 failed; the remainder were still in flight or abandoned when we looked.

For Standard quality, which is the bulk of it at 1,330 successful images, the midpoint was about 59 seconds. Nine in ten were done inside about 113 seconds. The slowest one percent took up to roughly 173 seconds.

High quality, 194 images, ran slower as you would expect: about 71 seconds at the midpoint, about 132 seconds for nine in ten, with a similar ceiling near 177 seconds.

The honest summary is: a minute, usually. Not instant. Long enough that you will look at your phone, short enough that leaving the tab is not worth it. If you want to know what each of those images costs rather than what it takes, that lives on the product facts page, which derives every figure from the code that charges you, so it cannot drift the way an article can.

Video is a different order of magnitude

337 video attempts in the same thirty days. 290 had completed successfully and 6 had failed by the time of the query, the rest still running or given up on.

The midpoint was around six minutes. The slow tenth stretched past twelve. That is not a defect and it is not something a faster queue fixes: generating a few seconds of moving image is genuinely a much larger job than generating one still frame, and every service doing it honestly is in the same range.

The practical consequence is a change in how you work rather than a number to endure. Images are a conversation with the model: try, look, adjust, try again. Video is not. You write the prompt properly the first time, start it, and go do something else. If you are new to that rhythm, how to write video prompts is worth reading before you spend six minutes finding out your description was ambiguous.

What the slow tail is actually made of

A midpoint of a minute and a ceiling of nearly three means something is occasionally much slower than usual, and it is worth saying what. Part of it is load at our provider, which we do not control and will not pretend to. Part of it is the work itself: larger resolutions, more reference images and longer clips all take longer, which is not a surprise but does mean the average hides real variation between a small square and a large widescreen frame.

We publish the tail rather than only the midpoint because the tail is the number that decides whether you trust the tool. A service that quotes a median and never mentions the worst case is quoting the case where nothing went wrong.

Waiting is not the same as risk

The reason a long wait is tolerable here is that it does not cost you anything when it goes wrong. A failed generation returns its coins to your wallet automatically; the mechanism is the same code path every time rather than a support ticket. We wrote that out in full in what a failed generation costs you, and the failure counts above are the reason we are comfortable pointing at it: 3 image failures and 6 video failures across a month is the regime that rule was designed for.

The part we cannot promise

None of this is a service level. There is no uptime commitment behind these numbers, no contract, and no guarantee that the month you generate in looks like the month we measured. Providers have bad days, and when ours does, our tail gets longer and you feel it.

What we can commit to is the honest version of the story: the figures here came from production, they include the failures, and we would rather publish a boring minute than an exciting adjective. If you want to see where your own attempts land, the image studio and the video studio both show you the clock while you wait.