news

Higgsfield's 10-minute AI video claim appears in an OpenAI customer story, with no sample and no availability detail

The jump from roughly one minute to over ten is attributed to Higgsfield's CEO, and the only evidence offered is the company's own YouTube channel.

By Daniel Ochoa · 15 October 2026

The headline number circulating since October 13 is that Higgsfield AI can now produce video longer than ten minutes, up from a ceiling of around one minute. The source is not a Higgsfield product page or a changelog. It is a customer story published on OpenAI's own site, and the claim is attributed to Higgsfield CEO and co-founder Alex Mashrabov, who says users "can create videos over 10 minutes long". There is no sample clip, no statement of which accounts or plans have it, and no measurement of the consistency problem the post says was the hard part.

The post credits three OpenAI models for different pieces of the workflow: GPT-6 Astra for length, GPT-Image-2.5 for precise visuals, and GPT-Live as a future path toward filmmaking directed by speech. Mashrabov describes voice-led creation as a direction the company wants to go, not something a customer can use today.

GPT-6 Astra coordinates the production, it does not draw the frames

Astra is a reasoning and tool-use model. In the account given, Higgsfield's creative agents use it to produce shot variations, call external tools and handle complex 3D modelling work, which is what keeps characters, environments and props from drifting as a piece gets longer. Nothing in the post identifies the video model that actually renders pixels, and that omission matters, because the longest native generations on the market are still measured in seconds. A ten-minute output is an assembly of many generations, held together by orchestration, rather than one pass of a video model.

That fits what Higgsfield sells. Its API offers more than 50 video and image models billed per generation, so the platform's quality on any given shot is mostly the quality of whichever third-party model is behind it. Per-generation cost is the figure that decides whether a ten-minute assembly is practical, and on our radar that varies widely by platform for the same model.

Continuity across cuts is the part that gets harder as length grows, and it is the part the customer story asserts rather than demonstrates. In our own benchmark we test it directly with multi-shot prompts, where a model has to keep a subject and a setting stable through a scene change.

The evidence cited is Higgsfield's own instructional videos

The example given is Higgsfield's YouTube channel, which Mashrabov says has almost half a million subscribers and teaches creatives how to use AI models and workflows. The team built a workflow that generates instructional videos over ten minutes long, replacing what the post describes as considerable manual effort. That is a real production change for Higgsfield's marketing output, and it is also the friendliest possible test case: tutorial footage tolerates cuts, screen content and narration in a way that a ten-minute narrative scene does not.

What is missing is anything a prospective buyer could check. There is no stated availability for the longer-video workflow outside Higgsfield's internal use, no per-account limit, no render time, and no published comparison against the one-minute baseline it replaces. OpenAI's page is a marketing case study about its own models, and it reads like one. Until Higgsfield publishes the workflow, the model list behind it, or a ten-minute clip anyone can inspect frame by frame, the number to carry forward is the claim and its source, not a capability.

Sources: OpenAI.

Daniel Ochoa: Covers model launches, shutdowns and pricing changes as they happen. Reads deprecation notices for a living so you do not have to.