GPT-6 Astra early tests: 15-min Mac app, 18-hour Pokemon run, city-scale 3D builds
量子位 · wechat · 2026-09-04
OpenAI's GPT-6 (Astra) is rolling out to paid ChatGPT tiers, and early testers have pushed it far beyond chat:
- 3D & spatial reasoning: Built a 3D iPod-style Codex client in Blender in 15 minutes; reconstructed full homes from Zillow photos with promo videos; built out Manhattan street by street; Higgsfield used it for actor blocking and shot planning on a film set, calling its spatial reasoning "world-class."
- Games: Beat Pokemon FireRed in 18h12m (vs 96h for GPT-5.6; GPT-5.5 never finished after 218h), using only screenshots; one-prompt Fall Guys-style and 10v10 Halo-style multiplayer games; Matt Shumer's Unreal Engine agents, each driven by Astra, spontaneously started socializing overnight.
- Fundamentals: Three.js ship-in-a-bottle with real-time physics plus its own music and video, SVG peacock, full music arrangement in Ableton.
Caveat: these "one-prompt" demos rely on model + Codex/MCP/Blender/Unity toolchains + long runs, and token burn is heavy—Terminal-Bench 4.0 tasks average 8M tokens ($17) each, with long projects potentially costing hundreds to thousands of dollars.
Related event: OpenAI launches flagship GPT-6 Astra to rave reviews and pricing debates(190 posts)→
More from Models
- Opus 5.5 tops OpenRouter spending in 20 of 29 task categories — mmurph · 2026-10-21
- Meta team builds open-source 30B LLM from scratch in months, hails 'second era' — TimDarcet · 2026-10-21
- DFlash drafter for Muse Glimmer-30B: 5 layers, blockwise 16-token spec decoding — TimDarcet · 2026-10-21
- EvoPress quantization shrinks 30B model from 56GB to 19GB at near-bf16 accuracy — TimDarcet · 2026-10-21
- Radar chart: Muse Glimmer-30B benchmarked against Gemma4-31B and Qwen3.6-27B — TimDarcet · 2026-10-21
- Privacy evals: Muse Glimmer-30B leaks less than Qwen3.6-27B on CIMemories — TimDarcet · 2026-10-21