Higgsfield pushes AI video past 10 minutes

OpenAI News says Higgsfield has pushed its AI video limit from about one minute to more than 10 minutes with GPT-6 Astra, and the startup is already using the stretch to build longer instructional videos for its own channel.

The jump is not just length.

Alex Mashrabov, CEO and co-founder of Higgsfield, told OpenAI News that consistency is the hard problem across longer clips, and he said Higgsfield’s creative agents now call external tools for 3D modeling work to hold shots together. The team points to its YouTube channel as a test case, which Mashrabov said has almost half a million subscribers and previously required considerable manual effort to produce. Now the team says it generates instructional videos over 10 minutes with an AI workflow, and Mashrabov linked the same capability to a personal project supporting rural schools in Kazakhstan where visual understanding is used to spot topics students struggle with.

For small businesses, the pitch shifts to still images. Mashrabov said GPT-Image-2.5 is the first model of its kind that can reliably render text and infographics on top of imagery, which matters when a promo needs readable pricing, menus or charts without a large production budget. OpenAI News frames that precision as the difference between a pretty picture and usable collateral.

Inside Higgsfield, the workflow change is organizational. The company builds with about 200 creatives and 100 engineers, per OpenAI News, and Mashrabov said Codex gives the creative side the power of a really strong software engineer. Creatives turn ideas into prototypes in Codex, work that previously needed heavy front-end and back-end engineering to reach production. Mashrabov said that handoff is now lighter because GPT-6 Astra can ship production-ready code, which he called never possible before.

What Higgsfield sells outside OpenAI’s post is broader than the four models named there. Its site describes an AI-native creative suite to create images, video and voice from prompts or references, with editing, upscaling and agent-automated workflows on web and mobile.

The source does not price longer video or quantify consistency. There are no benchmarks for drift, no latency figures for agentic tool calls, and no before-and-after cost for that YouTube workflow.

The next promise is voice to movie. Mashrabov said Higgsfield wants creators to describe an idea out loud and watch it render in real time, and he described OpenAI’s voice model as best in class for understanding and regular communication. Planning, tool calling, image generation and voice are the pieces he sees converging. No timeline or demo was shared, and OpenAI News did not say what the voice-to-video system would cost to run or how rights and likeness will be handled when speech becomes the director.