Breaking

Higgsfield Uses GPT-6 Astra to Build a 3D Oval Office in Blender

September 4, 2026 at 01:21 EDT

Foundation Models
Media Generation
AI Agents

Higgsfield used OpenAI's newly released GPT-6 Astra to generate scene code for a three-dimensional Oval Office, then built the location in Blender and rendered it with Cycles on Higgsfield Supercomputer. The company said a set-design description produced the scene code, after which Higgsfield assembled the location model and ran the full pipeline in its agentic workspace.

September 3, 2026 · Higgsfield

A 3D Oval Office, generated as scene code

GPT-6 Astra wrote the scene from a set-design brief. Higgsfield built the location in Blender and rendered it with Cycles on Supercomputer.

The pipeline

Describe set

→

Astra scene code

→

Build in Blender

→

Render with Cycles

Computer-use time per task

On OSWorld 2.0, GPT-6 Astra finished tasks in about 40 minutes, versus about 75 for GPT-5.6 Sol (about 47% less time).

~40 min

GPT-6 Astra

~75 min

GPT-5.6 Sol

Each block = 5 minutes. Shorter column is faster.

72.6%

OSWorld 2.0 score

~2x

Faster computer use

50%

Fewer manual fixes (Playco)

$10 / $50

API per 1M tokens

What landed

Scene code into Blender can be edited like any mesh, which developers say is more repeatable than prompt-only images.

Still open

A native Blender file was not released, the Supercomputer credit cost was not disclosed, and testers still prefer Claude for some visual 3D assets.

Continue reading

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