Higgsfield pushes AI video past the 10-minute mark

AI video has been stuck in the one-minute trap for what feels like forever. Try pushing past that, and faces morph, backgrounds glitch, and continuity completely falls apart.

Creative startup Higgsfield AI just blew past that ceiling. By integrating OpenAI's GPT-6 Astra, the platform has extended video generation to over 10 minutes while keeping characters and environments consistent.

Calling 3D tools under the hood

Higgsfield CEO Alex Mashrabov says the breakthrough came from changing how AI tackles the consistency problem. Instead of forcing a generative video model to remember everything, Higgsfield's creative agents now use GPT-6 Astra to call external 3D modeling tools for complex manual work.

That hybrid approach keeps scenes grounded across minutes of footage. Higgsfield even proved it on its own YouTube channel, which has nearly 500,000 subscribers, using the new workflow to generate full 10-minute instructional videos that previously took heavy manual labor.

The startup also plugged in GPT-Image-2.5 to solve another common frustration: illegible text. Mashrabov calls it the first model that reliably renders sharp typography and infographics directly onto images, giving budget-strapped small businesses a way to create polished promotional material.

Inside the company, the workflow looks just as different. Higgsfield employs roughly 200 creatives and 100 engineers. While the creative team previously used Codex just for rough prototyping, Astra is now generating production-ready code directly, cutting out the traditional engineering bottleneck.

The next step on their roadmap is voice-to-movie using GPT-Live. The goal is to let creators simply talk through their concepts and watch scenes render in real time.

Why it matters

Crossing the 10-minute mark transforms AI video from a novelty for 5-second b-roll clips into a real tool for long-form storytelling, tutorials, and full-length fashion campaigns. A solo creator can now produce commercial-grade work that previously required an entire studio.

We think Higgsfield's real genius here is relying on tool calling rather than waiting for raw model scaling to magically fix spatial consistency. Letting the LLM drive deterministic 3D software plays to the strengths of both worlds.

That said, chaining multiple model calls and external 3D pipelines is going to be computationally expensive, and real-time voice filmmaking will demand near-zero latency that current infrastructure might struggle to support reliably at scale.