GPT-6 Astra team member admits launch issues: code slop and excessive confirmations

yanndubs · x · 2026-09-04

Astra team member yanndubs shared first-hand launch notes on the new model: it's much smarter (evals saturating), more aligned, and better at computer use — including building a house in Blender. Known issues: too much code slop, and over-frequent confirmation requests from overcorrecting for caution. Practical warning: Astra follows instructions so strictly it executes unwanted ones buried in old skills, so clean those up before upgrading.

Related event: GPT-6 Astra Team Member Highlights Gains, Admits More Code Slop(4 posts)→

Original post →

More from coding & agent

coding & agent channel →