Matt Shumer's GPT-6 Astra review: Manager Loop sustains a week-long Manhattan build in Unreal
mattshumer_ · x · 2026-09-04
Matt Shumer published a 19-minute hands-on review of GPT-6 Astra. Key points:
- Daily driver: Plain-English replies, strong intent understanding and ambiguity handling; medium reasoning for everyday work, Ultra for big experiments. Claude still wins on visual taste and design assets.
- Manager Loop: The biggest surprise—a coordinator keeps the project moving while an implementer works in a separate Codex session, spawning sub-agents as needed. He used it to build a simulated civilization and make real progress on a GTA-like New York City.
- Manhattan build: Astra constructed a Manhattan world in Unreal Engine street by street over a week—work that would take humans months. Finding the setup that allowed such long runs took considerable iteration.
- Caveats: Long-running autonomy isn't solved; Astra can get absorbed in details and ambitious runs need careful coordination or they asymptote. It's the first model where he feels the right prompting can get it there.
- Cost: His large experiments consumed enormous token amounts.
More from coding & agent
- Bot Mesh: a social network where AI agents talk, host pages and pay each other — Daniel_Farinax · 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
- ChatGPT Work Mode vs Codex: cloud tasks or local control, use both — blakesamic · 2026-09-04