Guides
Two Ten-Minute AI Video Claims in One Week: A Checklist for Judging Long-Form Generation
ยท RenderBob team
Higgsfield said on 13 October that its AI video now runs past ten minutes. VideoGen said on 20 October that Model 1 can assemble ten-minute videos. The methods may differ.
Higgsfield said on 13 October that its AI video now runs past ten minutes, and VideoGen said on 20 October that Model 1 can assemble ten-minute videos. The headline number matches. The descriptions do not.
What each page says
Higgsfield's OpenAI customer story says videos that once topped out around one minute can now exceed ten minutes, using OpenAI's GPT-6 Astra, with agents handling shot variations to keep consistency. VideoGen says Model 1 plans screenplay, dialogue, shots, and pacing from a prompt and up to 16 reference images, and assembles longer videos from multiple generated shots. Both are company statements. Neither page links an independent test.
Three ways a ten-minute video gets made
A single long generation would mean one model holding a scene for ten minutes. Neither page describes that. The clip ceilings we have been tracking still sit near 20 seconds for FLUX 3 Video and 30 seconds for Kling 4.0 and Seedance 2.5.
Many clips stitched by an agent is closer to both descriptions. The system generates shots and assembles them. That is an editing pipeline as well as a model, and continuity, cost, and review time move with the number of shots.
Mostly static or repeated imagery can also run long. Higgsfield's story points at instructional video on its YouTube channel. Limited motion makes a long runtime easier, and it tells you less about a dramatic scene.
Questions to ask any vendor
- Is a ten-minute output generated in one pass or assembled?
- What is the resolution and frame rate?
- How many human review steps are in the pipeline, and who does them?
- What is the cost per finished minute, including regenerations?
- What drifts: faces, wardrobe, locations, voice?
- Do outputs carry provenance data, and does it survive export?
- Who owns the outputs, and what happens to your references?
A test you can run
Pick one character, one location, and one prop. Ask for a three-minute piece that returns to all three at the start, middle, and end, then score drift on each. Record how many generations you needed and how long review took. That is your real cost per usable minute. Our per-second price check covers the generation side of that sum. RenderBob lets you choose local or cloud for each clip and keep the references with the studio.
What to watch
Independent long-form tests, published pricing, and whether either vendor shows unedited output with the human steps listed.
RenderBob 3 and this story
RenderBob 3 lays a long piece out as shots on a timeline, so you can see whether the film is one pass or many takes. Each generated shot keeps its prompt.
More from the blog
- Four Voice Releases in Three Weeks: Eleven v4, Grok Voice 3.0, Fish Audio Drama 3 and Voxtral 3 Compared
Four voice models have launched since 28 September: Eleven v4, Grok Voice 3.0, Fish Audio Drama 3, and Voxtral 3. Their headline numbers measure different steps.
- MCP Gateways for Image and Video Models: What fal and Sunra Offer, and What to Check Before Connecting an Assistant
fal's hosted MCP server and Sunra's MCP page let an assistant run image and video models. The connection is also a new route for keys, prompts, and reference files.