The Festivals Want Finished Films. The Tools Make Seconds.

The biggest single prize in this piece went to a gaucho and a sheep. Listen to Meh, by the Argentine team Purga Studio — a gaucho whose days keep being interrupted by a sheep that will not stay off his roof — took first place and $500,000 at Higgsfield's festival, whose winners were announced on November 1. It is a finished film: a complete story, told from start to finish. The tools underneath this whole craft make four to fifteen seconds at a time. Everything interesting about AI film festivals lives in the gap between those two facts.

If you only remember one sentence, make it this one: the festivals are not judging the tools. They are judging finished films, and the rules are built to make sure a person actually made one. Everything below is that sentence, proved festival by festival — followed by the tool guides, which explain why the rules are shaped the way they are.

Rule set one: Runway's AI Festival — the open door

Runway's festival (it calls itself AIF) is the one with its rule book fully public, so it is the right place to start. For its 2026 edition, the film rules said:

• Your film runs between 3 and 15 minutes.
• It must include video made with AI tools — the rules name tools like Runway's own (its Gen-4.5 and Act-Two among them) as the reference point, which is the rule book's way of saying other makers' tools are allowed too. You are not required to use Runway to enter Runway's festival.
• Entering is free, and you upload the film to a free Runway account, or send a YouTube or Vimeo link — no other kind of link is accepted.
• With the film, you send a short written description of the AI techniques you used. No description, no entry. This is the rule that matters most, and you will see versions of it everywhere below.
• Entries ran from January 28 to April 27, 2026.

Judging is four scores, each marked from 1 to 10 by every judge, so the best possible total from one judge is 40: the quality of the whole piece put together; how well the story and its point hold together; originality; and the AI work itself — how inventively it is used, how skilfully, and how well it is woven in. Read those four again. Only one of them is about the tools. The other three are the same questions any film festival asks.

The prizes, from the same rule book: the top film takes $50,000 in cash plus a million Runway credits (the company values those at about $10,000); second takes $15,000; third $10,000; two further films take $1,000 each and five more take $500 each. Ten winning films in all.

Rule set two: Higgsfield — the closed shop, on purpose

Higgsfield runs the opposite shape of contest: the films are made inside its own platform, and its own pages publish the results. Its festival, whose winners were announced on November 1, paid out $1 million in total: $500,000 for first place, $200,000 for second, $100,000 for third, $100,000 for the audience's choice, and ten further prizes of $10,000.

Who decides: four judges from the film world — Edwin Catmull, Phedon Papamichael, Paul W. S. Anderson and Alex Proyas — picked first to third and the ten extra prizes, marking films on four things: directing and storytelling; the look of the work; sound and editing; and originality of the idea. The audience prize went to the film with the most valid votes — and only people on a paid plan could vote. Both of those sentences are from Higgsfield's own announcement of the results, and the second deserves its weight: the big money was judged, and the one voted prize was voted on by paying customers.

The winners, from Higgsfield's festival page: first, Listen to Meh, by the Argentine team Purga Studio — a gaucho whose days keep being interrupted by a sheep that will not stay off his roof. Second, Fukushima. Third, The Field Agent. The audience picked The Tortoise and the Hare.

And the company's own counts for the contest, printed here as Higgsfield's figures and nobody else's: 67,223 projects started, 5,000 finished films handed in, 176 countries represented, 55,628 people taking part. Now the sums — ours, and only arithmetic: roughly one project in thirteen became a finished film, and 14 films out of 5,000 took home money. That is the most honest picture of this craft going.

Rule set three: the African AI Film Festival — "central, or it doesn't count"

The newest festival on this page holds its first edition in Hollywood on November 21, and its own questions page gives the clearest one-line test anywhere: the AI work must be central to the film's making — used to create or add to the picture, the animation, the voices, the music or the writing. A film that only used AI to tidy up its colours does not qualify. Its words, not ours.

Its other rules: you must say which tools you used when you enter; shorts and animation run 1 to 10 minutes, documentaries 5 to 20; every film needs an African connection — a lead maker of African or African-diaspora heritage, a story centred on African stories, languages, history or culture, or an African collaboration; and entry was free for everyone up to September 30, and stays free for Africa-based makers and for students and newcomers.

The pattern, and the one that has no rules because it has no contest

Put the three side by side and the shared skeleton shows. A length limit. A complete film. A written note naming your tools. Judges marking story first. And then the differences, which are the interesting part: Runway lets you use anyone's tools, Higgsfield keeps you inside its own, and the African festival adds a test of what the tools were for.

Tribeca sits off to one side of all this, and it is worth saying exactly how, because it is often described loosely. Tribeca runs no AI contest and takes no AI entries through a rule book. It programmes work. In June 2025 it announced a year-long project with OpenAI: two chosen filmmakers, given money, guidance and training with OpenAI's tools, making live-action shorts to be shown for the first time at the 2026 Tribeca Festival, alongside a public conversation about the work. That followed its 2024 programme of five shorts made by past Tribeca filmmakers given early use of OpenAI's Sora. Invitation, not entry form — the door exists, but you cannot knock on it.

Now the tools, from their own guides

Here is why every festival above caps films at minutes, not hours — and why "a short written description of the techniques" is not paperwork but the whole craft.

Google's Veo 3.1 — Google's own model page says one go produces a clip of 4, 6 or 8 seconds, up to four clips at a time, as tall phone-shaped video or wide screen-shaped video, at everyday sharpness up to the very sharp 4K setting, with sound made alongside the picture. You can start from words alone, from a still picture, or from a first and last picture that the clip travels between. You can hand it example pictures to keep a person or a thing looking the same, and you can stretch a clip you have made further. That is the whole menu, in Google's own words, and it is a menu of seconds.

Kling's VIDEO 3.0 — Kling's own guide gives one go at 3 to 15 seconds, your choice of length inside that range, with sound made with the picture, including speech in five languages (Chinese, English, Japanese, Korean and Spanish). Its two headline tricks, again in its own guide's terms: it can plan several shots inside a single go — cutting from one view to another without you stitching the pieces — and it can lock onto a person or object you show it, so the same face stays the same face as the camera moves. Kling prices the work by the second: its guide lists a second of sharp 1080p footage with sound at 12 credits, and 8 credits without sound; at the softer 720p size, 9 credits with sound and 6 without.

OpenAI's Sora — closed. This is the entry most out-of-date guides get wrong, so here it is from OpenAI's own help pages: the Sora website and apps were switched off on April 26, 2026, and the service for other builders was switched off on September 24, 2026. OpenAI tells past users how to download their old work before it is deleted, and it has announced no replacement. If a site offers you "Sora" today, it is not OpenAI's tool, whatever the page says.

Versus, in one breath (the desk's versus format, kept honest): Veo gives you shorter pieces — 8 seconds at most — in a wider range of sizes up to 4K. Kling gives you pieces nearly twice as long, up to 15 seconds, and lets one go contain several shots. Both make the sound with the picture. Neither makes you a film. A seven-minute festival short is somewhere between 30 and 100 separate goes at these lengths, before you count the goes you threw away — and holding a face, a room and a voice steady across all of them is the work the festivals' written-description rule is really asking about. Nikolay Shestak's answer to that problem, in the piece linked below, was to start every shot from a still picture and reuse one sound sample per character. That is what "craft" means on this desk: not the tool list. The tricks for surviving the tool list.