Higgsfield assembles AI videos over 10 minutes, but not as one shot

According to an OpenAI case study, Higgsfield uses GPT-6 Astra to produce instructional AI videos longer than 10 minutes. CEO Alex Mashrabov says the previous workflow topped out at roughly one minute.

Ten minutes comes from a production workflow, not one continuous shot

Higgsfield's creative agents generate shot variations and, with GPT-6 Astra, can call external tools for manual work including 3D modeling. The company uses this process for videos on its YouTube channel, which Mashrabov says has almost half a million subscribers. Higgsfield's public Explainer page separately advertises videos up to 10 minutes long.

OpenAI also connects the longer format with GPT-Image-2.5 for visual assets. The account does not describe one continuous ten-minute output from a video model. A more accurate reading is a production pipeline that plans, generates and assembles multiple parts into a finished instructional video.

The value lies in connecting steps, not merely adding minutes

For a small creative team, consistency of characters, style, text and voice across scenes matters more than a record-length render. An agent that preserves the plan and calls the right tool can absorb repetitive shot variations and parts of the 3D workload. A person can edit the result instead of manually carrying every step between tools.

Higgsfield brings together roughly 200 creatives and 100 engineers. The company says creatives use Codex to turn ideas into prototypes, while GPT-6 Astra reduces the work needed to put them into production. That organizational change is more interesting than the 10-minute figure: the line between designing a campaign and building internal software is getting thinner.

The case study omits price, speed and the cost of corrections

The source is an OpenAI customer story based mostly on claims from Higgsfield's CEO. It provides no benchmark, price, render time, resolution, failed-generation count or comparison with human post-production. It also does not say which plans or regions can access the entire workflow.

The voice-to-movie language around GPT-Live-1 describes a future ambition. Higgsfield wants spoken direction and real-time rendering of intent, but the source does not confirm a finished customer feature. Claims about consistency over longer videos deserve the same caution until complete exports and their editing process are visible.

An unedited export and production bill will reveal the real advance

The decisive evidence will be public ten-minute projects that disclose scene count, regenerations, time and credits. Equally important is the ability to revise one scene without breaking characters, voice or continuity elsewhere in the video.

If the workflow can repeatedly produce a long program with little intervention, it moves AI video from isolated clips toward a production system. If it requires a hidden army of editors, 10 minutes remains merely the duration of the final file.

Lilith's verdict

A ten-minute export is the finish tape, not footage of the whole race. Higgsfield wins only when it shows how many generations, credits and human edits stayed outside the frame.