Astra update: evals saturating, better CUA, but asks for confirmation too often
himanshustwts · x · 2026-09-04
Yann announces Astra is out: much smarter (evals are saturating), more aligned and trustworthy, with better CUA — demonstrated by a house Astra built in Blender. Known issues: too much code slop, and Astra asks for confirmation too often, which can feel lazy; the team overcorrected toward caution and will fix both next.
More from coding & agent
- Bot Mesh: a social network where AI agents talk, host pages and pay each other — Daniel_Farinax · 2026-09-04
- Codex release rust-v0.153.1 quietly adds API config for unreleased GPT-6-Astra model — github-actions[bot] · 2026-09-04
- One prompt, $2: GPT-6 Astra generates an entire Subway Surfers-style game — mhdfaran · 2026-09-04
- Playco turned one grey-box kart prototype into 3 Unity games with GPT-6 Astra — mhdfaran · 2026-09-04
- SPACE cuts agent LLM calls by 78.9% while raising success rate on long-horizon tasks — dair_ai · 2026-09-04
- GPT-6 Astra tops Zapier's AutomationBench at 41.4%, first model ever to clear 40% — sandersted · 2026-09-04