Creator uses GPT-6 Astra and the Blender MCP add-on to build walkable 3D scenes in Blender

• Model: gpt-6-astra. Standard API pricing is $10 per million input tokens and $50 per million output tokens. Fast mode runs about 2.5x faster at twice the price. Enterprise admins must switch it on per workspace (DataCamp).
• What Blender MCP can do: create and change objects and materials, run any Python code inside Blender, check the scene from several viewport angles, pull free CC0 assets from Poly Haven and Poly Pizza, and generate models with Hyper3D Rodin, Tripo or Hunyuan3D (GitHub).
• Requirements: Blender 3.0+, Python 3.10+ and the uv package manager. Listed clients include Claude Desktop, Claude Code, Codex, Cursor and VS Code (GitHub).

• How it is built: an add-on inside Blender runs a socket server. A separate MCP server passes tool calls between the AI client and that socket (GitHub). The model works mostly by writing Blender Python scripts. Ahuja describes this as using models' coding strength to make up for weak spatial reasoning (AI Engineer).
• The demo's method:

1. Generate a reference image.
2. Build the scene, sorted into named collections (Blender's folders for objects).
3. Set materials and lighting based on that image.

The demo then asks the model for a report on what it built (video, 01:46; 04:31).

• Reproducibility: none. There are no prompts verbatim, files, run times, costs or failure rates. The demo also contradicts itself on method: it credits Astra's on-screen computer use (02:00), but Blender MCP works through scripting.

• Astra leads in computer use: OpenAI reports 72.6% on OSWorld 2.0, against 65.7% for GPT-5.6 Sol, with about 47% less time per task (DataCamp). This is OpenAI's own figure.
• ARC-AGI-3 score: OpenAI's 99.9% used a costly custom test setup. ARC Prize's independent runs with the standard setup reportedly scored about 17–63% (DataCamp).
• "Beats learning Blender": the only support is the creator's own example. Its first scene could not be navigated until a fix was requested (video, 04:11).
• Reference-image reconstruction is a real feature: this matches the project's documentation (GitHub).

Blender MCP predates GPT-6 and was built mainly around Claude. Ahuja's 2025 talk showed scene generation, reference-image rebuilding and animation through Sketchfab, Poly Haven and Rodin (AI Engineer). CGWorld covered the add-on's updates in 2025, and Ahuja has also built an Ableton MCP (Cline blog). Astra's launch materials name FreeCAD and KiCad among the professional apps it can operate (Wikipedia). The Blender use is a community add-on, not an OpenAI feature.

• Security: arbitrary Python execution is a real risk. The project offers a "safe mode" that blocks file access and network calls. Telemetry is on by default but can be turned off (GitHub).
• Scale limits: complex operations may need to be broken into steps, and Poly Haven downloads freeze the Blender interface until they finish (GitHub).
• Model concerns: OpenAI says Astra crosses its "critical" cybersecurity threshold (Fortune). An anonymous source reported a reasoning method that hides some of the model's reasoning from monitoring (Wikipedia).
• Source bias: the video repeatedly promotes a paid "AI Profit Boardroom" (03:00).

• Good fit: fast concept scenes, mood pieces and early product or website mockups, where a "strong draft" (04:27) is enough.
• Not shown: production-grade geometry or ready-to-ship assets.
• Before deploying: turn on safe mode, save a copy of the Blender file before each run, check asset licences, and budget for Astra's output-token pricing.
• Lock-in: the bridge works with any listed client, so teams can compare Astra against Claude or Gemini rather than commit to one vendor.