Higgsfield's GPT-6 Astra workflow pushes AI video past 10 minutes

In OpenAI's customer story published October 13, 2026, Higgsfield CEO and co-founder Alex Mashrabov says videos on the platform "topped out at around one minute" before GPT-6 Astra and now run over 10 minutes. The same day, Higgsfield's own Supercomputer page and site links advertise an agent "powered by GPT-6 Astra," so the integration is live at the vendor, not just promised in a newsroom post.

How a 10-minute video gets held together

The hard part of long AI video is not raw duration; it is consistency across a 10-minute timeline. Mashrabov describes Higgsfield's creative agents handling repetitive production work such as generating shot variations, and calling external tools for 3D modeling tasks — an agentic capability of GPT-6 Astra that the CEO credits for keeping characters, sets, and style aligned between scenes.

The company's own YouTube channel, which Mashrabov says has close to half a million subscribers, is the first test bench. Higgsfield built a workflow that generates instructional videos over 10 minutes long there, replacing manual production effort. Duration claims in vendor stories are marketing until a team publishes long outputs on a channel people actually watch, and this one at least shows its own work.

Images, prototypes, and a voice roadmap

Two other details in the story matter for working creators. Higgsfield uses GPT-Image-2.5 for promotional assets, and Mashrabov calls it "the first model of its kind that can reliably render the text and infographics on top of the imagery" — the failure mode that has made AI images unusable for ads and pricing graphics.

Internally, around 200 creatives and 100 engineers build the product. Creatives prototype in Codex, and the CEO's claim that GPT-6 Astra "can ship production-ready code" is the kind of workflow shift many small teams are chasing, though the article offers no benchmark behind it. The forward-looking piece is GPT-Live: Mashrabov describes a "voice to movie" direction, with OpenAI's voice model as the enabler. The article frames this as a goal, not a feature you can use.

What is unconfirmed, and how to verify it yourself

Be precise about the evidence: the over-10-minute figure is a CEO statement in an OpenAI newsroom post, and neither Higgsfield's Supercomputer page nor its pricing page publishes a maximum clip duration or the plan tier required. There is no independent, measured test showing output coherence at 10 minutes — that gap is the real editorial question behind the headline.

Verification takes a few clicks. Open the linked GPT-6 Astra Supercomputer agent in Higgsfield's app, ask for a multi-scene video of three or more minutes, and inspect character consistency between cuts. On your plan, check where the per-clip cap, credits per minute, and queue behavior actually sit. If a result is shareable, the YouTube channel and contest submissions are public reference points.

Why LinkLoot tracks this

Video length is currently the frontier constraint creators hit first, and a 10-minute generation path at a mainstream suite changes what solo teams can deliver without a studio. The same article also lists GPT-Live as Higgsfield's next step, which is the signal to watch: voice-driven editing would move AI video from prompting to directing.

For tooling around agent-driven creative workflows, see LinkLoot's guide to AI agent tools.

Sources and verification status

1. OpenAI: Higgsfield takes AI video from 1 minute to over 10 — primary official source; contains CEO attribution and all duration and model claims.
2. Higgsfield Supercomputer — vendor source; independently confirms the GPT-6 Astra-powered agent and the image/video/audio product line, but does not publish a duration spec.

Unconfirmed: exact maximum video duration per plan, availability of the 10-minute workflow for all users, and any measured consistency evidence at that length. GPT-Live voice-to-video is announced as a roadmap direction, not a released feature.