Higgsfield has extended its AI video generation limit from 1 minute to over 10 minutes, integrating three new OpenAI models: GPT-6 Astra for extended video coherence, GPT-Image-2.5 for frame-level visual precision, and GPT-Live for real-time voice-directed filmmaking.
The GPT-Image-2.5 integration is the detail worth examining closely. Frame-level precision in AI video has been the hard problem, not length. Getting a model to maintain consistent lighting, character features, and scene continuity across 600-plus seconds is a different engineering challenge than generating a compelling 10-second clip. The original piece explains how Higgsfield is solving for both simultaneously.
GPT-Live points toward a workflow shift: directors speaking edits in real time rather than typing prompts and waiting. Whether this holds at scale, or collapses under latency and interpretation errors, is the open question. Read the full piece for the technical breakdown of how these three models interact inside Higgsfield's pipeline.
[READ ORIGINAL →]