Hands-on with GPT-6Astra driving Blender: 3 workflows, real costs and pitfalls
数字生命卡兹克 · wechat · 2026-09-08
A detailed Chinese tutorial on using GPT-6Astra (via Codex) to drive Blender, noting the hype has spiked Blender downloads. Setup requires ChatGPT desktop, the official Blender MCP server (not the community one), and the ComputerUse plugin.
Three workflows are compared:
- ComputerUse: the agent models like a human, click by click — it built a detailed Temple of Heaven hall in 4 hours, but consumed nearly half of the $200 Pro quota in one task.
- Blender MCP: faster and cheaper via the bpy API; the author generated a detailed motorcycle model plus a 10s assembly animation, though long tasks can time out and should be split.
- CLI Python scripts: functionally similar to MCP, little difference in output.
Key limitation: AI handles regular geometry (buildings, vehicles) well but fails at organic characters, which are better outsourced to Tripo AI then refined in Blender through the full pipeline of modeling, UVs, materials, rigging and animation. The author closes on creation becoming democratized — what differentiates people now is taste and judgment, not software skill.
More from coding & agent
- Ex-OpenAI engineer pushes back on Codex origin story: it began as an internal infra tool — dotey · 2026-09-08
- dembrandt: extract any website's design system into W3C tokens in one command — tom_doerr · 2026-09-08
- Sentdex: GLM 5.3 Flash-level capability would've been overkill a year ago — Sentdex · 2026-09-08
- Agents automate dirty high-leverage work — veterans benefit most, not idea guys — code_star · 2026-09-08
- MemContinuum: open-source long-term decision memory for Claude Code projects — ShotPorter · 2026-09-08
- facebook-marketplace-mcp: an MCP server that queries Facebook Marketplace via GraphQL and Chrome cookies — Signalman23 · 2026-09-08