Higgsfield Pushes AI Video Past 10 Minutes With GPT-6 Astra, Up From a 1-Minute Ceiling

OpenAI said Tuesday that AI video platform Higgsfield has lifted the maximum length of the videos it can generate from roughly one minute to more than ten minutes, and it is crediting the jump to GPT-6 Astra, the frontier model OpenAI released at the end of September.

Duration has quietly been the hardest number in AI video. Generating a striking ten-second clip has been table stakes for two years. Holding a character's face, a scene's lighting and a narrative's visual logic across minutes of footage is what breaks, because generation errors accumulate and the model drifts. Most consumer tools have capped output well below the length where that drift becomes unwatchable.

Higgsfield's approach is to hand the consistency problem to agents rather than to a single model call. GPT-6 Astra supports creative agents that can invoke external tools, including manual 3D modeling work, and the company says it uses that ability to keep characters, scenes and visual detail aligned across long sequences. The claim is plausible on its face: tool use is exactly what long-horizon coherence tends to require, whether the intermediate work is a rig, a layout file or a shot list.

The company's own YouTube channel, which has close to half a million subscribers, doubles as the demonstration. Higgsfield says its team has used the GPT-6 Astra workflow to produce instructional videos longer than ten minutes, replacing what had previously required a large amount of manual production work. That is a self-reported result, and quality across ten minutes of footage is not something an announcement can settle, but the channel gives outsiders something they can check.

The update is not just about length. Higgsfield has also integrated GPT-Image-2.5, which the company says can render readable text and infographics on top of images reliably, a persistent weak point in generative imagery. For small businesses producing marketing material, legible text on a flyer or a chart is often the difference between usable output and a discard.

Internally, Higgsfield uses OpenAI's Codex to accelerate product development. The company says it runs roughly 200 creative staff alongside 100 engineers, and that creatives can now turn ideas into working software prototypes themselves, reducing the amount of front-end and back-end engineering time each feature needs.

The next step the company has outlined is voice-driven creation built on GPT-Live-1: describe an idea out loud, have the system interpret the intent and generate video in real time. That would push the workflow from prompt-and-wait toward something closer to directing.

The broader signal is where frontier value is being claimed. A year ago, model launches carried the news cycle; now the measurable gains are showing up in application workflows, where a vendor's model is the engine but the differentiator is what a partner builds around it. For OpenAI, a partner shipping a tenfold jump in usable output length is a better advertisement for Astra than another benchmark table.