News
Higgsfield Says AI Video Now Runs Past 10 Minutes: What an OpenAI Customer Story Does and Doesn't Show
ยท RenderBob team
An OpenAI customer story from 13 October says Higgsfield can now produce AI video longer than 10 minutes using GPT-6 Astra. The claims are the company's, and no independent test has been published.
An OpenAI customer story from 13 October says Higgsfield can now produce AI video longer than 10 minutes using GPT-6 Astra. The claims come from the company and its CEO, and no independent test has been published.
What the page says
The story, filed under Startup on OpenAI's site, says Higgsfield used to top out at about one minute per video and can now go beyond 10 minutes. CEO Alex Mashrabov says GPT-6 Astra also powers creative agents that call outside tools, including 3D modelling, and handle repetitive jobs such as shot variations, which the company says helps keep longer videos consistent. The page also names GPT-Image-2.5, which Mashrabov says renders readable text and infographics, and a heading points at GPT-Live-1 only as a future direction for voice-driven "voice to movie" filmmaking. The body does not say GPT-Live-1 is available.
Higgsfield says its YouTube channel, which Mashrabov puts at almost half a million subscribers, uses AI to generate instructional videos over 10 minutes long. The page describes about 200 creatives and 100 engineers working on the product, and an early project using GPT-6 Astra's visual understanding to make tailored learning videos for rural schools in Kazakhstan.
What it does not say
The page gives no exact maximum length, only "over 10 minutes". It has no pricing, no regional availability, no shot-level examples we could assess, and no mention of provenance metadata or content credentials. It also does not say whether a 10-minute video is generated in one pass or assembled from many clips by agents. If agents are assembling clips, the job looks like editing. If one pass produces the ten minutes, the job looks like generation.
How to read "10 minutes"
Compare it with the clip-length ceilings we have been tracking. Wan 3.0 and Seedance 2.5 are reported at 30 seconds per clip. A 10-minute output built from agent-managed shots is a different kind of claim, closer to our two-lane view of editing, with generation in one lane and assembly in another. Check identity at minute eight, and again at the end. A clean opening minute does not answer that.
A fair test is to take one character, one location, and one prop, and ask for a video that returns to all three across ten minutes, then check identity drift at the start, middle, and end.
Instructional video is the first target
The examples given are instructional and fashion-campaign work, where scenes are repeatable and tolerance for drift is higher. Narrative drama is a different job, because continuity errors are visible. The claim that a single creator can run a high-end fashion campaign from start to finish is Mashrabov's.
What to watch
Independent tests of long-form consistency, published pricing, and whether longer outputs carry provenance data. Our Article 50 coverage explains why that last point matters in the EU.
More from the blog
- Disney Bids $20 Million for a License to Character.AI's IP: What the Stalking-Horse Deal Does and Doesn't Say
Character.AI filed for Chapter 11 on 12 October, and Disney Worldwide Services is the opening bidder at $20 million for a non-exclusive, perpetual licence. Higher bids can still come in.
- Nicolas Cage Won't Sign Amazon's AI Waiver: What Performers Are Being Asked to Agree To
Nicolas Cage says he refused to sign an AI waiver for Amazon's Spider-Noir and will probably not work with the studio again. The waiver's terms have not been published.