GPT-6 built a rocket in-game after 6.1M tokens and $367 in compute
Angaisb_ · x · 2026-09-16
The author gave GPT-6 Astra Ultra a single goal: build a rocket ship in The Universim. After 7h 50m, 6.1M tokens and $367.40, it succeeded.
- The model had a CLI built beforehand by a separate thread (2 hours to build and test), then controlled the game via that CLI and Computer Use, with subagents handling threats, CLI usage and industry.
- The author didn't expect it to work given how slow Astra is, but the model was smart enough to pause the game, plan its next moves, then execute.
Clarification: the original post didn't make clear the CLI was built beforehand.
Related event: GPT-6 Spends $367 and 8 Hours to Build Rocket in Game(4 posts)→
More from coding & agent
- Success Rates of Public MCP Servers Benchmark Against Merged Results — shensi · 2026-09-16
- Training Agents video series complete: 6 videos, 8 hours from SFT to RL environments — SergioPaniego · 2026-09-16
- How do you cap an AI agent's API spend across multiple vendors? — apyhubnico · 2026-09-16
- Sistava launches AI employee platform for business workflows, plans from $25/mo — Mahmoud_Zalt · 2026-09-16
- Multi-turn agent RL training at scale on HF Hub: 9,523 sandboxes in 14h, zero crashes — vanstriendaniel · 2026-09-16
- Does anyone actually use Codex ultra mode? Subagents just produce 'a mountain of slop' — wstone_bd · 2026-09-16