🛠 GPT-6 Astra beat Portal in a day
User @cozyblazex on X ran GPT-6 Astra in maximum reasoning mode through the original Portal: 23 hours, 42 minutes, and 34 seconds from start to credits according to the log. The model guided Chell via an MCP server and a SourcePauseTool patch — 3,210 portal_exec commands. At API pricing, the run would have cost $574 with caching and $4,400 without; in reality, it fit within a $200 Codex subscription.
🌍 A frontier model held a single goal in a 3D environment for nearly a day using only screenshots and telemetry — a rare public demonstration of long-horizon agency, and the open log sets a benchmark for the economics of such runs: input token caching cut the price by roughly 8x.
👤 The repository includes a walkthrough video, a sanitized log, and the full setup: controller, SourcePauseTool patch, and a Windows guide. Input token caching is the main factor in the cost-effectiveness of multi-hour agentic runs.
Source 1: https://x.com/cozyblazex/status/2096383114851533097
Source 2: https://github.com/cozyblaze/portal-agent