Breaking

Higgsfield says GPT-6 Astra leads motion video when run on its Supercomputer

September 5, 2026 at 01:23 EDT

Foundation Models
Media Generation

Higgsfield AI says OpenAI’s GPT-6 Astra now produces motion video that outperforms dedicated AI video models, with the clips generated end to end on the company’s Higgsfield Supercomputer. The claim is a product demonstration rather than a published third-party video benchmark, and it sits inside a broader early-access push that treats Astra as a planner, coder, and 3D operator rather than a native text-to-video generator.

Sep 2026 · Higgsfield AI × OpenAI

Astra directs the motion. Video models finish the shot.

Higgsfield says GPT-6 Astra now leads motion video when the full pipeline—scene code, Blender geometry, Seedance 2.5—runs on Higgsfield Supercomputer. The model is a planner and 3D operator, not a native video generator.

Production loop

1 · Astra

Map space, write scene code, lock blocking

→

2 · Blender / Cycles

Editable geometry and path-traced light

→

3 · Seedance 2.5

Final motion pass, clips up to ~30s

Computer-use time, drawn to scale

OSWorld 2.0 latency simulations: GPT-6 Astra finishes in about 40 minutes per task versus about 75 minutes for GPT-5.6 Sol — roughly 47% less time.

~75 min

GPT-5.6 Sol

~40 min

GPT-6 Astra

Shorter column, faster agent: about 47% less time at a higher 72.6% OSWorld score.

What landed

Code-based motion, landmark rebuilds, and in-frame blocking look unusually coherent when Astra plans and Seedance renders.

Open questions

No independent video leaderboard backs the “best video model” line. Joints, multi-round consistency, and token cost still need work.

“Astra … executes our most complex creative workflows while using up to 20% fewer tokens than other models we've tested.”

Alex Mashrabov, CEO and co-founder, Higgsfield AI · API list price $10 / $50 per million input / output tokens

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.