OpenAI's 10-minute Higgsfield claim is about the agent, not the video model

The title on Higgsfield's YouTube channel says it all happened in one sitting: "GPT-6 Astra + Higgsfield MCP Made This ENTIRE Video in One Chat." The upload runs 11 minutes and 7 seconds, went up on Sunday 6 September, and has passed 311,000 views. The description underneath tells a longer story. The presenter clips were generated from the host's own face and voice, the motion graphics were animated from existing templates in Adobe After Effects, and the whole thing was assembled in DaVinci Resolve.

Keep that video in mind when you read the line now traveling around AI video circles. On Tuesday 13 October, OpenAI published a customer story about Higgsfield under the headline "Higgsfield takes AI video from 1 minute to over 10." Its opening paragraph reads: "Before GPT-6 Astra, those videos topped out at around one minute. Now, says CEO and co-founder Alex Mashrabov, users can create videos over 10 minutes long." Within hours, aggregator accounts on X and at least one AI news site were reporting that Higgsfield had raised its generation cap tenfold. That is not what OpenAI's own page goes on to describe.

What the OpenAI page actually says

Read past the opening and the claim narrows quickly. The first sentence of that quote sits outside quote marks, so it is OpenAI's narration, and no direct quote from Mashrabov anywhere in the story mentions minutes. When the page gets to evidence, it points at one place: Higgsfield's own YouTube channel, which teaches creatives how to use AI models and workflows. According to OpenAI, that content used to take considerable manual effort, "But with GPT-6 Astra, the team has built a workflow that uses AI to generate instructional videos over 10 minutes long."

So the 10-minute figure describes an in-house workflow that assembles tutorials for Higgsfield's channel. The story names no user-facing feature, shows no sample, quotes no price and cites no customer who has made a 10-minute film this way. OpenAI doesn't say which channel videos it means, either. The "One Chat" upload is the closest match we found by title and length, and we read its description and chapter list rather than watching the full video. The channel subscriber count holds up: the YouTube Data API showed 476,000 subscribers on Wednesday, in line with the "almost half a million" Mashrabov gave OpenAI.

Astra plans the shoot, other models render it

GPT-6 Astra is OpenAI's frontier language and agent model, sold in ChatGPT and to developers through its API (the programming interface other apps plug into) as gpt-6-astra, and it doesn't output video. In the customer story it does what agent models do: Higgsfield's creative agents take on repetitive production tasks like making variations of a shot, and with Astra they can "call external tools to handle complex, manual work in 3D modeling." Mashrabov says that helps with "maintaining consistency across longer videos," which the page asserts and never shows.

The footage itself still comes from video models, and Higgsfield routes several of them through its platform. Its product pages haven't moved. Cinema Studio, the film tool, describes each generation as a scene of up to 30 seconds, with up to 50 references for consistency and an Extend control to carry a shot forward or back. The Marketing Studio help page caps clips at 15 seconds per generation. Nothing we found on Higgsfield's site offers a 10-minute single generation, and OpenAI's story never says which video models made the footage.

The one-minute baseline doesn't hold either

The "before" half of the headline has its own problem. Higgsfield's channel was publishing long AI work well before anyone was talking about Astra. The YouTube Data API lists Hell Grind, which Higgsfield bills as an AI feature, at 95 minutes, posted on Tuesday 4 August, and a 1 hour 55 minute upload called Cully Hill Boys on 10 August. The channel has also carried dozens of tutorials running 10 to 37 minutes since April.

Those were human-directed productions, and the Hell Grind production diary cites a crew of 15 and 10 million credits. Read generously, then, the honest version of OpenAI's claim is that Higgsfield could not automate a long video from start to finish before Astra, and says its agents now can. That is a genuinely interesting claim about orchestration, and a much smaller one than "AI video goes from one minute to ten."

Who gains from the line

Customer stories are marketing, and both companies have a stake in this one. Higgsfield launched inside ChatGPT in August, ships an MCP connector for ChatGPT and Claude, and has now appeared in at least three OpenAI customer stories: one on cinematic social videos on 21 January, one on shipping features in a day on 21 September, and this one. OpenAI is selling Astra, and Higgsfield is selling its ChatGPT integration. Higgsfield reported a $400 million Series B at a $5.4 billion valuation in August, alongside $700 million in annualized revenue, a figure the company reports itself.

Higgsfield, for its part, hasn't amplified the story at all. A search of its X accounts over the past seven days for OpenAI, Astra, GPT-6 or minutes returned no posts. Its week on X was about Katana, which runs on Anthropic's Claude Motion, and a Wan 3.0 promotion. The rest of the OpenAI page is in the same register: Mashrabov calls GPT-Image-2.5 "the first model of its kind that can reliably render the text and infographics on top of the imagery," says "GPT-6 Astra can ship production-ready code, which was never possible before," and describes a "voice to movie" future built on GPT-Live-1 with no product, date or price attached. An early effort with rural schools in Kazakhstan gets one paragraph and no detail.

Where the real opening is

Strip the headline away and something useful is left. The long video in this story was never a single render. It was a plan broken into short scenes, each one generated, checked and stitched, with a person finishing the cut. That is the same workflow working AI filmmakers already run, and the part Higgsfield says Astra now handles is the tedious middle: shot lists, variations, consistency passes and tool calls.

In practice, for anyone making work this week, the takeaway is that orchestration is the skill to build, and it doesn't belong to one vendor:

• Plan longer pieces as runs of 15 to 30 second scenes, because that is still the unit every major platform renders.
• Try a language model with tool access as your shot planner and continuity checker, whether that is Astra, Claude or another model.
• Budget real finishing time in an editor, since Higgsfield's own "One Chat" video lists After Effects and DaVinci Resolve in its credits.

The question worth putting to Higgsfield is simple: is the 10-minute workflow something a paying user can run today, what is it called, and how much of the channel's output still passes through a human editor? Until it answers, the claim describes Higgsfield's own studio, not your next render.

Timeline

• Date: Wed 21 Jan 2026. What happened: First OpenAI customer story on Higgsfield, about cinematic social videos; Source: OpenAI (title and date only)
• Date: Tue 4 Aug 2026. What happened: Hell Grind, a 95-minute AI feature, posted to Higgsfield's YouTube; Source: YouTube Data API
• Date: Wed 12 Aug 2026. What happened: Higgsfield launches inside ChatGPT; Source: Higgsfield YouTube
• Date: Mon 17 Aug 2026. What happened: $400M Series B at $5.4B valuation, $700M annualized revenue (self-reported); Source: PR Newswire, TechCrunch
• Date: Fri 4 Sep 2026. What happened: First Higgsfield GPT-6 Astra tutorial on YouTube; Source: YouTube Data API
• Date: Sun 6 Sep 2026. What happened: "GPT-6 Astra + Higgsfield MCP Made This ENTIRE Video in One Chat" (11:07) posted; Source: YouTube Data API
• Date: Mon 21 Sep 2026. What happened: Second OpenAI story: Higgsfield ships features in a day with Astra; Source: OpenAI
• Date: Tue 13 Oct 2026. What happened: Third OpenAI story: "from 1 minute to over 10"; Source: OpenAI
• Date: Tue 13 Oct 2026. What happened: AI news site and X aggregators report a raised generation cap; Source: AI In Context, X

Sources and what is unconfirmed: OpenAI customer stories of 13 October and 21 September 2026 (read in full) and its GPT-6 Astra page; Higgsfield's Cinema Studio page and Marketing Studio help page (read 14 October); YouTube Data API pulls of Higgsfield's channel and uploads (14 October); X recent search of Higgsfield's accounts (7 days); PR Newswire and TechCrunch for the Series B. We did not watch the "One Chat" video in full, OpenAI does not name the 10-minute videos, and which video models render the footage is unconfirmed. We have not yet put our questions to Higgsfield or OpenAI. Higgsfield's revenue figure is self-reported.

Your Voice

Quoted from public posts on X and credited by handle. If one of these is yours and you want it removed, message @the__generated on X and it comes down.

Got something to say? Tag @the__generated on X