Updated October 15, 2026 · 6 min read
GPT-6 Astra takes Higgsfield's AI video from 1 minute to over 10 — here's what's claimed and what isn't proven
On October 14, 2026, OpenAI announced that Higgsfield AI — a creative video platform — can now generate videos over ten minutes long, up from a previous ceiling of about one minute, using GPT-6 Astra as the underlying model. The same announcement bundles GPT-Image-2.5 for precise visuals and a first step toward voice-directed filmmaking with GPT-Live. The headline number is a big deal for AI video. The caveats deserve equal billing.
What was announced
Higgsfield AI's mission is professional-grade video and image creation for anyone with an idea. Before GPT-6 Astra, its videos topped out at around one minute. Now, says CEO and co-founder Alex Mashrabov, users can create videos over ten minutes long.
The proof point offered is Higgsfield's own YouTube channel, which has almost 500,000 subscribers and teaches creatives how to use AI models and workflows. That content used to require considerable manual effort. With GPT-6 Astra, the team built a workflow that generates instructional videos longer than ten minutes — and per the announcement, a single creator can now execute a high-end fashion campaign from start to finish.
How GPT-6 Astra makes longer video possible
Two mechanisms are described in the announcement:
- Extended generation itself. GPT-6 Astra powers the video engine that pushes past the one-minute ceiling. Duration has quietly been AI video's hard constraint — most consumer-facing models produce clips measured in seconds, forcing creators to stitch dozens of short generations together.
- Tool-calling creative agents. Higgsfield's agents take on repetitive production tasks, such as generating multiple variations of a shot. With GPT-6 Astra, they can also call external tools to handle complex manual work in 3D modeling. The company credits this with attacking one of creative AI's hardest problems: maintaining character, scene, and visual-detail consistency across a long video.
Mashrabov also gave a personal example: he supports rural schools in Kazakhstan, and an early effort uses GPT-6 Astra's visual understanding to identify topics students struggle with and generate videos tailored to what they need to learn.
GPT-Image-2.5 and the road to voice filmmaking
The announcement is a three-model bundle, not just a video update:
- GPT-Image-2.5 is positioned for precise visuals. Higgsfield calls it state of the art and "the first model of its kind that can reliably render the text and infographics on top of the imagery" — aimed at small businesses making promotional material without large production budgets.
- GPT-Live is framed as "a path to voice filmmaking" — creators describe what they want out loud rather than through a technical interface. The next step, per Higgsfield, is GPT-Live-1 voice-to-video: a creator describes an idea and sees it take shape in real time. Note the phrasing — "a path" concedes the destination has not been reached yet.
- Codex handles the software side: Higgsfield runs about 200 creatives alongside 100 engineers, and creatives use Codex to turn ideas into software prototypes without heavy front-end and back-engineering involvement.
What's claimed vs. what's proven
This site's pattern applies here too: separate the headline from the evidence. The announcement states the capability but does not publish:
- No benchmarks or comparisons against competing video models (Runway, Pika, Luma, Sora).
- No pricing, no availability dates, no resolution or frame-rate specs.
- No independent full-length generations — until outside testers produce them, "over ten minutes" is a stated maximum, not a demonstrated typical result.
- Long-form coherence — whether characters and scenes actually stay consistent across ten minutes — is asserted, not demonstrated. It is historically the hardest problem in AI video, and duration without continuity is not the same achievement.
One more structural point: Higgsfield is a middle-layer company building on OpenAI's foundation models rather than training its own. The terms of that relationship — licensing costs, exclusivity — are undisclosed, which matters for whether rivals can simply match the duration gains with the same underlying models.
Frequently asked questions
How long can Higgsfield AI videos be now?
Over ten minutes, up from about one minute, per OpenAI's October 14, 2026 announcement. Higgsfield's own YouTube channel (almost 500,000 subscribers) already uses a workflow generating instructional videos longer than ten minutes. It is a vendor-stated capability — no independent benchmarks yet.
What does GPT-6 Astra actually do in Higgsfield's pipeline?
It powers the extended video generation that breaks the one-minute ceiling, and its creative agents call external tools for complex manual work like 3D modeling — which Higgsfield credits with solving the consistency-across-long-video problem.
What are GPT-Image-2.5 and GPT-Live?
GPT-Image-2.5 renders readable text and infographics on imagery for small-business promotional content. GPT-Live is a voice interface framed as "a path to voice filmmaking," with GPT-Live-1 voice-to-video as the stated next step.
Why does the ten-minute length matter?
Duration is the hard constraint that keeps AI video in short-form territory. A ten-minute ceiling, if it holds up in independent use, moves AI video toward narrative-length output and lets a single creator run a full campaign end to end.
What should I be skeptical about?
The announcement includes no benchmarks, pricing, or independent full-length samples, and long-form coherence is asserted rather than demonstrated. Treat "over ten minutes" as a stated maximum until outside testers verify it.
Related reading
For how Astra's agent capabilities work more broadly, see computer use and Daybreak gating and what GPT-6 Astra is. For the pattern of headline-vs-honest numbers, see what the benchmark scores really mean.