FULL STORY

GPT-6 Astra Hands-On: Stunning 3D, Shaky Instruction-Following

Early community tests of GPT-6 Astra showcase impressive 3D and game creation, but users consistently flag weak instruction-following as the model's key shortcoming.

2026-09-08 ~ 2026-09-11 · 2 episodes · 21 posts

Episode 1 · GPT-6 Astra Showcases: Community Tests Full 3D Creation Pipelines (2026-09-08, 15 posts)

After the release of GPT-6 Astra, the community posted a wave of hands-on 3D results. X user @sidahuj compiled a series covering games, animation, film previz, scene reconstruction, product design and cloth simulation, with most projects finished in 1-4 hours. Overall, Astra shows strong 3D workflow capability, though multiple authors note quality stability depends on persistent prompting; natural-language-driven workflows (including Japanese) are significantly lowering the barrier to 3D creation.

Confirmed

  • A user generated a complete post-apocalyptic game with shooting and HUD in one shot from just a few material/model links, a worldview description and a core idea.
  • lofihashbloom shared the first game demo "The Legend of Astra", combining Unity CLI, Blender MCP, Tripo and Make-… (item 14 in sidahuj's series).
  • @jumperz showed possibly the first online game built with Astra: single-player or online PvP, built with Astra + Blender in about 4 hours; still early but "could be really good" with polish.
  • @TheMoonMidas rounded up 8 interactive demos: juggling robots, playing Portal, Super Smash Bros. on a coffee table, and more.
  • @nemumusitocha built a castle via a Blender MCP workflow, from creation to rendered video in about 1 hour.
  • sidahuj tested npaka123's workflow to reconstruct a classroom scene from a reference photo via Blender CLI and Blender MCP.
  • @Stefan3DAI was most struck by Astra's spatial understanding: with Blender MCP and Tripo P2.0 it generated optimized models, assembled them (a house on a turtle's back), rigged characters and produced animation end-to-end.
  • @mrblackstudio demoed headphone design ideas with Astra + Blender MCP, alongside castle, game and classroom cases.
  • @simonsmith used Astra Ultra + Blender MCP to build skin and animation models for a biotech project; quality became presentable only after explicit demands, noting "you have to keep pushing it".
  • Japanese user @keisukeoyama finished modeling with just two instructions in Astra high, calling the quality "moving".
  • @yossy222VRC directed Astra via Blender MCP to build a DJ booth purely in Japanese natural language (incomplete due to token exhaustion), never touching Blender manually.
  • Another user had Astra via MCP generate realistic 3D models such as a bat, iteratively refined.
  • @higgsfieldai showed Astra via Higgsfield MCP building a robot model ready for animation in a single Blender session, suitable for film previz, claiming "AGI moment is not far".
  • @minchoi compiled 10 examples of Astra generating 3D games, Blender scenes, human anatomy and real-world object models from near-zero input.
  • Developer majidmanzarpour demonstrated cloth physics simulation in three.js using Astra with Blender and a custom workflow, offering to publish a tutorial.

Why it matters

  • These cases show that Astra combined with Blender MCP, Unity CLI, Tripo, three.js and other tools already covers a fairly complete 3D pipeline - modeling, scene reconstruction, assembly, rigging, animation, cloth simulation, rendering and finished games - mostly within 1-4 hours, sharply cutting creation costs; natural-language-driven cases (including Japanese) further lower the barrier.
  • Multiple authors report unstable output quality requiring repeated pushing, and some projects were cut short by token exhaustion - a key variable in assessing real-world utility.

Episode 2 · Hands-on with GPT-6 Astra: Stunning 3D Games, but Reliability Issues Hold It Back (2026-09-09, 6 posts)

After GPT-6 Astra's release, Peter Yang and several other users ran deep hands-on tests. The consensus: "both stunning and maddening." 3D modeling and game development are widely seen as its signature strengths, but instruction-following and reliability issues keep it from becoming an everyday workhorse model.

Confirmed

  • Peter Yang set GPT-6 Astra as his default model from launch, calling it "both amazing and frustrating." Strengths include 3D modeling and game development (calling Blender to build 3D models, assembling them into playable games with Godot, with no need to learn Blender), high-level thinking and strategy, and computer/browser control.
  • Two major weaknesses: ① it sometimes only describes what it plans to do instead of just doing it; ② it cannot reliably follow user-defined AI usage rules (skills).
  • @majidmanzarpour, replying to Peter Yang, shared a similar experience: Blender 3D modeling was excellent, but coding tasks still required heavy prompting to complete.
  • @NextTechnology6361 deeply tested GPT-6 Astra with Blender MCP for 3D modeling and found the real experience far from the launch demo: results degrade markedly when asking Astra to directly generate or modify geometry, and using a set of vintage fire truck reference images with Tripo3D produced unsatisfactory results.
  • @葬AI tested the full GPT-6 Astra 3D modeling-and-printing pipeline: non-human objects were reproduced with high fidelity, but human figures failed. Conclusion: a big step toward productization, but 3D capability remains rough—"low-hanging fruit." Subjectively there was no leap in base intelligence; benchmark scores rose but nobody is talking about intelligence gains—productization is the real story of this update.

Why it matters

  • Jensen Huang said GPT-6 Astra means "AGI has arrived," but users' hands-on results show a clear gap, indicating there is still distance between launch demos and everyday usability.
  • The model dramatically lowers the barrier to 3D creation (modeling, assembling games, and hooking up 3D printing without learning Blender), representing a productization-driven rather than purely intelligence-driven direction, with significant implications for the creator tool ecosystem.