WHAT NEW IN X10.30 FRI

X 上的 AI 合集档案 · 22 个合集 · 60,637 条

gpt-6VIDEO

Astra 写码、Claude Code 审码的 Codex 流程

作者把 Claude Code 嵌进 Codex,让 Astra 写限流器、Claude 挑错,实测 Astra 跑分强但仍有盲点。

Sanchit monga
@sanchitmonga222026.09.14收录原推发布于 2026.09.06
Sanchit monga @sanchitmonga22原推发布于 2026.09.06
在 X 上打开原推

GPT-6 Astra is here so I tried it as my primary engineer. Astra writes the code. Claude Code reviews it. Inside Codex. One prompt. Back in July OpenAI shipped a plugin that puts Codex inside Claude Code. I built the reverse: Claude Code inside Codex. Astra builds a rate limiter, hands it to claude, claude gives out findings that astra can then accept/reject, fix and ship. Honest take after a day on it: Astra is a monster. Beats Fable 5.1 on DeepSWE, Terminal-Bench, Terminal-Bench Science by 12 points, doubles every published Claude score on ARC-AGI 3, and does it at about half the cost per task. That is why it is the engineer here. But Claude still tops the independent Coding Agent Index, and in this run it caught things Astra wrote and walked right past. Different lab, different blind spots.

x.com/sanchitmonga22/status/2096372075439235580采集于 2026.09.14 16:10