Higgsfield AI Video Jumps Past 10 Minutes With GPT-6 Astra Agents

A tenfold jump in running time

Higgsfield AI, the creator-focused generative media platform, says its users can now produce videos longer than 10 minutes. Until now, its output topped out at about one minute. The claim comes from co-founder and CEO Alex Mashrabov, who credits OpenAI's GPT-6 Astra model for the change.1 OpenAI published the news as a customer story on October 13, 2026. The post also covered two other OpenAI models: GPT-Image-2.5 for more precise still images, and GPT-Live for a future of voice-directed filmmaking.36

The headline number is easy to repeat, but it needs a closer look. The coverage agrees on the basic facts. It differs on how much those facts prove, and on what a "10-minute video" actually means in this case.

How the longer videos get made

OpenAI's account does not describe a single model that renders ten minutes of footage in one pass. It describes a system of agents. Higgsfield's creative agents already handled repetitive production work, such as generating several variations of the same shot. GPT-6 Astra lets those agents call external tools for harder, normally manual jobs like 3D modeling.1 Mashrabov says that tool use helps with one of the hardest problems in creative AI: keeping things consistent as a video gets longer.2

That framing matters. Making AI video longer is mainly a continuity problem. Characters, sets, objects and visual style all have to hold steady across scenes, or viewers notice.3 The Chinese outlet AIBase described the agent layer as something like a self-organizing assistant director, there to stop faces from changing between shots.5 That describes the goal well. Nobody has published independent evidence yet that the problem is solved.

Some outlets oversimplified the story. One summary said the software can now make much longer videos "in one go."4 AIBase called it a partnership that raised the length limit for AI-generated video in general.5 The OpenAI post supports a narrower reading: Higgsfield built an orchestration workflow, run by a stronger reasoning model, that assembles longer pieces.1 Lapaas Voice gave the most careful reading. It noted that a longer video is not just a matter of raising a duration setting, and that the announcement does not say whether every user, model or generation mode can reach the new length.3

The proof point: tutorials, not cinema

The main real-world example is Higgsfield's own YouTube channel. Mashrabov says it has almost half a million subscribers and teaches creators how to use AI models and workflows. Producing those videos used to take a lot of manual work. The team now has a GPT-6 Astra workflow that generates instructional videos longer than 10 minutes.1

The choice of example is telling. Tutorials are a forgiving format. They depend on clear explanation, diagrams and steady pacing, not on long dramatic scenes that test character continuity. Lapaas Voice pointed out that no controlled comparison of the old and new workflows was published, with no hours saved and no measure of how much human editing is still needed.3 In my reading, Higgsfield has shown that long-form explainers work. Long-form narrative has not been shown.

Mashrabov also described a personal project. He supports rural schools in Kazakhstan, and an early effort there uses GPT-6 Astra's visual understanding to spot topics students are struggling with and generate videos aimed at those gaps.1 It is one of the more concrete examples of what longer, personalized AI video could be for, though it is described as early.

Image generation and voice: the multimodal stack

Length is only part of the announcement. Mashrabov says GPT-Image-2.5 can put readable text and accurate infographics into images. He presents that as useful for small businesses that can't afford conventional production.2 Readable text has long been a weak spot for image generators, so this is a meaningful claim. As with the video claim, no accuracy benchmarks came with it.3

The voice part looks furthest out. Mashrabov wants to reach what he calls "voice to movie": a creator describes an idea out loud and the AI interprets it and renders the film in real time.1 He sees OpenAI's progress on voice, planning, image generation and tool use as steps toward that.2 Every outlet that looked closely agreed this is a plan, not a product. Blockchain.News and TokenPost both called voice filmmaking a direction of development rather than something finished.67 Taken together, the three pieces describe a full multimodal pipeline: a reasoning model to plan and coordinate, an image model for exact visuals, and a real-time voice model as the interface. But only the first of these has a shipped result attached.

The context: an aggregator that once gave up on long-form

Higgsfield's history makes this announcement more interesting. Former Snap AI chief Mashrabov launched the platform in spring 2025 with a focus on cinematic camera movement. At launch, the company said it was working on character and scene consistency over longer sequences. Later in 2025, SaaStr reported that Higgsfield had moved away from building its own long, coherent narrative videos and instead aggregated top third-party video models, many of them Chinese. At that point reviewers found clips capped at five seconds and weak consistency over time.

That approach clearly paid off commercially. Reports in January 2026 said an $80 million extension brought Higgsfield's Series A to $130 million at a $1.3 billion valuation, with annualized revenue reaching $200 million within nine months of launch. The company's tools pull in outside models such as Sora 2, Kling and Minimax, alongside its own Soul image model. In September 2026 it said its developer API offered more than 50 video and image models.6

Seen this way, the 10-minute claim is not a sign that Higgsfield has built a better base video model. It shows the company's orchestration layer improving: the agents that string many short generations into something coherent. Those agents have also gotten better because the reasoning model behind them did. The company already pointed this way in May 2026 with its Supercomputer agent, which plans, picks models, generates and stitches clips. One early user reported outputs sometimes running several minutes, plus a company claim that a 23-minute pilot was made in 96 hours. The same user said coherence dropped noticeably past about 60 seconds, with drift between segments.14 That one-minute point matches the earlier ceiling Higgsfield now says it has passed.1 So the specific improvement on offer is that GPT-6 Astra's tool use, including 3D modeling, reduces the drift that used to set in after a minute.

What's still unknown

Reporters agree on what is missing. There are no technical benchmarks, no pricing changes and no rollout details. It is also unclear which customers, workflows or content types can use the longer-video capability.67 It isn't stated whether the feature is generally available or limited to certain workflows. Nothing independent measures character continuity, generation time or cost per minute.3 Day-to-day reliability is another question. In early October, one user said a 15-second render on the platform was stuck processing for 13 hours, which other users blamed partly on free-tier queues.8 An anecdote proves nothing, but it shows how far marketing claims can sit from everyday use.

Why it matters anyway

Even with those caveats, the announcement shows where AI video generation is going. Competition is moving away from which model makes the best five-second clip and toward which system can plan, keep track of context and coordinate many models and tools into a finished piece. For text-to-video, that makes the reasoning model as important as the diffusion model. Mashrabov's stated goal is to let smaller creative teams do more ambitious work, even letting a single creator run a high-end fashion campaign from start to finish.1

I read the announcement as real progress described in sweeping terms. Higgsfield appears to have pushed AI video past the one-minute mark for structured formats like tutorials, using agents that lean on a stronger reasoning model. Whether that holds up for drama, for paying users at scale, and for the real-time voice workflow it has promised is something the company has not yet shown.