Higgsfield says GPT-6 Astra takes anime film from brief to export

• An Astra-directed anime short moved from a sword-fight brief to Blender previs, Seedance 2.5 shots, and a final edit, according to the sword-fight demo.
• 3D scene construction sits upstream of the renders in Higgsfield’s workflow: the floor-plan demo starts with scene code, while the museum sequence adds cast and shot-list planning.
• The posts target editable project files rather than flattened output, with the After Effects demo describing transparent, reusable comps and the Figma demo ending in an editable reconstruction.
• Higgsfield has also put Astra beside Fable 5.1 in a ghost-story comparison, a paper-craft Kabuki comparison, and a side-by-side video.
• An independent Blender reconstruction used photogrammetry-derived textures and two roughly 45-minute critic loops, as Bilawal Sidhu’s project and his follow-up describe.

OpenAI’s launch post says Astra is rolling out from a limited set of organizations to paid ChatGPT tiers, the API, Azure, and Bedrock, while a post from chrisfirst called the model official. Its system card describes safeguards around autonomous capabilities. An OpenAI game-building walkthrough uses TypeScript, Vite, and Three.js to keep direct control of geometry, materials, and lighting; Higgsfield’s production announcement says its team is applying computer use to a first movie.

The four stated steps in the sword-fight demo are:

1. Develop a sword-fight story.
2. Block the previs in Blender.
3. Generate shots with Seedance 2.5.
4. Cut the shots into a final version.

Higgsfield says the process retained the characters and location across the project. The post does not include a shot count, prompt log, edit decision list, or continuity measurement.

The accompanying production thread shows desktop control used for visual-reference research, asset generation, and timeline editing.

The recurring production state in these examples is a Blender scene, with generated video arriving after spatial layout and blocking.

• A floor plan becomes scene code, then a Blender location model, then a Cycles render in the blueprint-to-architecture post.
• A museum sequence maps the location, cast, and shot list before blocking stays in the viewport and Seedance 2.5 renders each setup, according to the spatial-reasoning demo.
• A separate long-take sample links phone movement to the Blender viewport and travels through seven rooms without a cut in the handheld-camera post.

Higgsfield’s claims extend beyond generating clips. Its After Effects demonstration says Astra analyzes a reference video, recreates motion to its timing, and places each element in a separate editable composition with transparency for reuse.

The Figma project is described as a layer-by-layer rebuild from one reference image. The final file remains editable, according to the Figma reconstruction.

Other app-specific examples add two operational details: the Illustrator demo says the agent built a native document and routed around a frozen picker through Finder to export it, while the DaVinci Resolve post says it matched a reference grade from flat Apple Log 2 footage in four minutes.

Higgsfield’s Astra versus Fable 5.1 material is a creative A/B series, not a published benchmark. The three prompts shown are:

• A hand-drawn Japanese ghost story, with both outputs rendered in Seedance 2.5 in the ghost-story post.
• A paper-craft Kabuki scene, also rendered in Seedance 2.5 in the Kabuki post.
• A future-city side-by-side identified only as an evaluation in the comparison post.

The posts provide neither shared prompt text nor seeds, attempts, costs, or a scoring method. A Reddit creator repeated the format with the same ink-wash myth and Seedance 2.5, asking viewers to judge storytelling, motion, composition, and consistency in the community comparison.

Bilawal Sidhu gave Astra an old 3D scan of a family drawing room and reported that it rebuilt the space in Blender without downloading assets, instead using textures from the photogrammetry scan and procedural shaders.

Sidhu then ran two loops of about 45 minutes each, using a critic to compare the work against multi-view imagery before leaving the run overnight, as his iteration update explains. The same evidence collection includes a separate link-only visual clip, whose post text supplies no model or workflow detail.