GPT-6 hands-on: 41-second keyboard setup, Blender pirate ship, and AI video editing put to the test

卡尔的AI沃茨 · wechat · 2026-09-06

A Chinese blogger ran extensive Codex-powered tests of GPT-6's computer-use and workflow abilities, finding major gains though its design output still trails Claude.

Reasoning UI gripe: The reasoning-effort control is a slider whose leftmost position silently downgrades to GPT-5.6; the author initially mistook it for a three-step selector.
Peripherals: GPT-6 configured three custom keys on a $10 keyboard in 41 seconds on the first try (GPT-5.6 failed repeatedly), then set 14 functions on a Logitech MX Anywhere 3s across browser, Feishu, and Codex with almost no manual fixes.
3D: Built a Blender pirate ship in under 10 minutes with no downloaded models, generated a 15-second keyframed camera-move video a la Your Name, and turned three Mario levels into a spherical 3D game nobody has beaten.
AI editing: After three iterations, GPT-6 produced a publishable vertical talking-head video, solving letterboxed horizontal footage with a two-layer layout. A full editing prompt is included.
Design comparison: On a skill-gallery webpage, Claude Fable 5.1 still beats GPT-6 on design language, notably with a three-image merged cover that avoids case/skill cover duplication.

Original post →

GPT-6 hands-on: 41-second keyboard setup, Blender pirate ship, and AI video editing put to the test

More from coding & agent