GPT-6 Astra hands-on: 2x faster, stunning 3D frontends, coding on par with Claude
数字生命卡兹克 · wechat · 2026-09-05
A detailed first-day review of GPT-6 Astra by Chinese blogger 数字生命卡兹克:
- Availability: Rolling out to all subscribers with 100% quota (vs Fable 5's 50% cap); much faster than GPT-5.6 Sol.
- Speed: Tasks that took 3+ hours now finish far quicker; large system reviews 10 minutes; computer/browser use roughly 2x faster, though interaction still trails Grok and Gemini.
- Frontend: One-prompt generation of detailed interactive 3D scenes (lunar rover, explorable house models with material/wireframe/floorplan modes) — a massive aesthetic leap.
- Coding: The same project-scan prompt surfaced many previously missed performance issues, batch-fixed in 2 hours. New experimental cross-window notes let Codex retrieve old messages/tool results, reducing summary loss.
- Practical advice: Simplify AGENT.md rules for smarter models — the author shares his trimmed config and a prompt for Codex to self-audit AGENTS.md/Skills; over-engineering and excessive testing notably reduced.
- Weak spots: Blender control via Computer Use still shaky; Chinese writing still lacks "restraint", though less robotic than GPT-5.6 Sol.
Related event: GPT-6 Astra rolls out globally to rave early reviews(95 posts)→
More from coding & agent
- Open-source MCP server lets Claude offload small tasks to opencode's free models — BridgeCrafty · 2026-09-05
- Coding agents reviewing each other led to recursion hell — a simpler outcome-driven workflow — boringmarketer · 2026-09-05
- Simplify your codebase workflow: a business-outcome prompt beats agent cross-reviews — boringmarketer · 2026-09-05
- This team let a model named Fable run onboarding interviews for every new AI model — hrishioa · 2026-09-05
- Cursor's 1-second webview delay traced to VS Code 1.99 bug, fix landed upstream in 1.102 — max__drake · 2026-09-05
- dotmd author: command-palette launch (shift cmd p) builds muscle memory fast — max__drake · 2026-09-05