GPT-6 Astra

OpenAI's architectural walkthrough shows Astra using Blender's Python API, bpy, to create geometry, materials, lights and cameras. It reviewed preview renders and revised the scene before delivery. Later versions used Poly Haven environment assets and textures, so this example combines procedural modeling with existing resources. The Blender scene was also taken into Unreal Engine 5 for an interactive walkthrough.

In our September 6 Codex/Blender test, a second donut was rebuilt from generated design and multi-view references, then checked with orthographic renders. Compared with the first version, it had a thicker body, shorter rounded sprinkles and thinner icing with rounded drips. Both editable Blender projects were retained. Inspect the comparison and play the synchronized video below; the captions link the test settings and limitations.

V1 · Direct modeling

V2 · Reference-guided revision

Synchronized turntables: V1 on the left and V2 on the right · 8 seconds · No audio

• The 8-second comparison keeps the camera, animation, lights, plate, background and render settings fixed; both source animations are 1080p at 24 fps. Blender 4.5.9 LTS rendered with Cycles at 32 samples on an RTX 4060 using OptiX and denoising.
• The second scene contains 286 objects, including 250 sprinkles. Mesh checks found zero non-manifold edges in the dough and evaluated icing; this is not a complete self-intersection or printability test.
• The reference views had residual perspective and inconsistent scale. Geometry and material edits plus review produced the improvement; this was not automatic exact image reconstruction or an isolated reference-input experiment.
• The result remains stylized, especially the bread microtexture. One iterative workflow with human direction cannot establish a model ranking, a one-shot success rate or superiority over another provider.

In its September 5 Sydney Opera House post, Higgsfield says Astra generated the building as code, Blender produced editable geometry, and Cycles supplied path-traced lighting on Higgsfield Supercomputer. Its September 3 Oval Office example describes the same scene-code, Blender and Cycles pipeline. These are Higgsfield's accounts of its demonstrations; they do not establish measured reconstruction accuracy or provide a guarantee about every generated mesh.

The examples extend beyond a single campaign. Tom Krcha describes turning an old steam-train drawing into 3,295 editable Blender objects. At the September 6 observation, that post showed about 1.60 million views and Higgsfield's Oval Office post about 1.16 million. Those counters indicate substantial attention to individual demonstrations, not unique viewers, a measured growth rate or independently verified modeling quality. Krcha's object count and completion time remain his reported results.

In a September 5 roundup, Will Hu (@willlhhh) argues that the notable shift is an agent operating Blender, Unreal Engine and a browser to produce projects people can edit, explore or play. His follow-up frames complex desktop software as a promising use of computer use, while calling these early demonstrations rather than established production reliability. This is his interpretation. OpenAI's documented Blender workflow combines Python scripts with visual inspection, so screen interaction, APIs and MCP should not be treated as mutually exclusive routes.

Higgsfield's September 3 seven-room shot credits Astra with mapping the space and keeping the staging in the viewport, while Seedance 2.5 renders the final video. This is a combined spatial-planning and video workflow. It should not be read as Astra directly outputting the finished video or as evidence that a downloadable 3D model was delivered.

OpenAI reports 95.9% geometric overlap for Astra with tools on BenchCAD, which asks models to reconstruct 3D objects from multiple rendered views by generating CAD code. The launch comparison lists 83.3% for GPT-5.6 Sol and 84.3% for Claude Fable 5.1. This is a publisher-reported, task-specific result; it does not measure animation quality, production topology, architectural accuracy or success on every Blender request.

Use a tool-enabled environment with access to Blender, then start with one room or object. OpenAI's walkthrough offers a useful sequence: review the shape, inspect renders, request changes, and only then expand the scene or export it to another engine. A successful first prompt can still involve many tool calls and revisions.

• Specify the deliverable: an editable Blender scene and the script that builds it, plus a few preview views.
• Give reference images, dimensions where relevant, the intended use and a manageable level of detail.
• Inspect separate objects, hidden surfaces and the result of one requested edit before accepting the asset.
• For a game, test the exported asset in the target engine; for printing, check the mesh in a slicer. A polished render alone does not answer either question.
• Track modeling time, rendering time, usage and manual correction separately. Higgsfield's infrastructure is part of its examples, not a requirement established for all Astra workflows.

In a September 6 report, @hylarucoder preferred Astra to Fable 5.1 for their backend work. They described porting sqlc's Go implementation toward Node.js and getting through most cases using roughly half their quota, clearer Chinese conversations and documentation, and less intrusive Electron debugging through computer use. They also reported fewer corrections to intent and implementation. This is one developer's experience, not a controlled comparison: the quota window, full test results and matched Fable configuration were not supplied. Their experience with fewer refusals or account restrictions does not establish either provider's enforcement rate.

@hylarucoder recommends medium for daily work because high did not feel sufficiently better to justify the additional usage in their tasks. Treat that as a starting hypothesis for your own tests. OpenAI supports low, medium, high, xhigh and max; its migration guidance says to preserve your effective effort setting, or start with low when moving from none or minimal. Compare task completion, corrections, latency and consumption before changing your default.

OpenAI now lists GPT-6 Pro, powered by Astra, in Chat for Pro $100, Pro $200, Business and Enterprise plans; Enterprise also depends on workspace model permissions. Plus includes Astra in ChatGPT Work and Codex, while Chat, Work, Codex and API allowances remain separate. Astra requires Codex CLI 0.153.0 or newer and the latest desktop app, so check the exact product surface instead of treating one plan label as universal access.

OpenAI's September 22 family launch keeps Astra as its best model across the board, while Sol targets complex coding and professional work at $2/$10 and Luna targets focused, high-volume work at $0.10/$0.50 per million input/output tokens. Start lower when your own acceptance checks pass; escalate to Astra when the harder task benefits justify its price and access boundary.

The documented API model ID is gpt-6-astra. It accepts text and image input, returns text, supports low through max reasoning effort, and provides a 1,050,000-token context window with up to 128,000 output tokens. Base rates are $10 input, $1 cached input, $12.50 cache write, and $50 output per million tokens.

• Requests above 272,000 input tokens use 2x input and cache rates plus 1.5x output rates for the full request.
• Batch and Flex cost 50% of standard rates; Fast mode costs 2x the applicable rate.
• Responses API tools include web and file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, and tool search.

OpenAI calls Astra its most capable model, while independent September 3 testing shows strong coding-agent gains but a mixed value story on a broader intelligence index because the model costs 2.5 times GPT-5.6 Sol at current promotional rates. OpenAI also classifies Astra at the Critical cyber threshold, so organizations should validate authorization boundaries, monitoring, and review workflows alongside task quality.