Thirteen questions. 971 takes.

The short version

FLUX.3 is better at reasoning than at arithmetic, and better at building a world than at cutting one.

It tracks cause and effect further than we expected. A seven-step Rube Goldberg machine holds all the way to the plate, an off-screen noise reliably arrives just before the on-screen reaction, and asking for twenty seconds doesn't stretch a five-second idea — it writes a longer script, with a turn the short take never had. Name a period and you get the frame shape, the title cards and the cut rhythm of that decade, not a filter over a modern shot.

It cannot count. Ten coins knocked off a stack land as thirteen; a scorebug reads perfectly and then increments wrong — 132 of 145 requested strings rendered exactly while almost anything that had to tick or total went astray. It also won't take direction on pace: you can contradict a format's setting, its subject and its lighting, but not the speed of its cutting, and cuts land no closer to a musical beat than chance. Pinned keyframes get the same treatment — first and last are targets, the ones in between are suggestions, and ten pins land less often than four.

Two things matter less than the internet says. How you format a prompt is nearly irrelevant — eight of ten writing styles came back interchangeable as long as the prompt still said who and where. And length has a ceiling: the model reads every word out to 4,700 characters, but the film stops improving at about 1,400. One thing matters more: the language you write in. Write the same scene in Cantonese and it recasts the actors and rebuilds the street — though only where the scene has a street.

And when you reroll, know what moves. The face, the lens and often the ending change; the location, the light and the look do not. A longer prompt steadies the picture without steadying the plot — which is the whole argument for drafting several takes cheaply before committing to one.

How it was run

Every take is a single generation, straight out of the model — nothing cut, graded or retouched. Each experiment fixed a scene and moved one variable across it, so a difference on screen is the variable and not the roll. Where a claim is a rate (“27 of 32”), it was counted by watching the takes, and the denominator is printed with it. Where a result could be luck, it was tested against a shuffled baseline and the report says so.

Nothing here is a benchmark score. It is thirteen production questions, answered with enough takes that the answer isn't one lucky render — and the full board is public, so you can disagree with the reading and check the footage yourself.