Skip to content
AI Primer
release

Higgsfield adds GPT-6.2 Astra to its creative agent tools

Higgsfield added GPT-6.2 Astra to its creative agent tools. Demonstrations include a video built from a 23,000-line Python program and 3D reconstructions from archival footage and historical book scans.

3 min read
Higgsfield adds GPT-6.2 Astra to its creative agent tools
Higgsfield adds GPT-6.2 Astra to its creative agent tools

TL;DR

A 573-part cinema camera and a stealth platformer inside a painting feature in Higgsfield’s showcase. OpenAI singles out gaming and 3D design as areas of improvement, while Higgsfield’s ChatGPT workflow library puts Blender workflows alongside ad-production templates.

Python-built scenes

Procedural scene-building is the standout: Higgsfield says its White House interior was created without any hand-modeled 3D assets. The company attributes the city film below to a 23,000-line Python program.

In a Houdini and Blender demonstration, Higgsfield describes three production steps:

  1. Build the Vellum and OpenVDB network in Python.
  2. Replace keyframed splines with center-of-mass physics.
  3. Layer 243 sound elements and render on Higgsfield GPUs.

Archival footage

Higgsfield’s archival examples turn an early film and a 1969 Apollo launch clip into 3D spaces with freely movable viewpoints.

Scientific book scans

Skin deformation and dioramas

  1. Character animation: Higgsfield claims real-time skin deformation on complex anatomy, retaining pore detail as multiple eyes emerge from a face.
  2. Walkable scenes: another demonstration lets the viewer enter a living 3D diorama.

Hands-on tests

AmirMushich asked Astra and Fable 5.2 to create and animate the same model in Blender on an RTX 4070 laptop. The reported first-pass usage was:

  • Astra: 3.5 million tokens, including 26,000 output tokens, for approximately $3.12.
  • Fable 5.2: 21.2 million tokens, including 236,000 output tokens, for approximately $19. About 98% of its tokens were cached context.

AmirMushich preferred Astra’s modeling, but slightly favored Fable’s rendering and movement logic. After a second Astra iteration, the two sessions had consumed 31.3 million tokens in total; rendering improved while animation deteriorated.

Every’s team, which tested the models for several weeks, reported three limitations in its early assessment:

  • Writing: answers became excessively terse compared with Astra 6, including when longer explanations were appropriate.
  • Coding: benchmark performance was similar to Astra 6, but end-to-end builds frequently stopped too early.
  • Design: strong default instincts produced a repetitive “house style.”

Access and credits

Higgsfield’s connection guide describes installation through ChatGPT’s Plugins Directory, followed by connecting a Higgsfield account. An active paid Higgsfield subscription is required, with no API key needed.

The MCP credit policy limits unlimited access and free generations to the Higgsfield website. Generations through MCP and other automated tools deduct credits at standard rates.

Further reading

Discussion across the web

Where this story is being discussed, in original context.

On X· 4 threads
TL;DR4 posts
Python-built scenes2 posts
Scientific book scans2 posts
Hands-on tests2 posts
Share on X