GPT-6 Astra autonomously completes Portal: Costs, tools, and AGI questions

OpenAI's GPT-6 Astra just played through Valve's Portal completely on its own, using 3,336 tool calls, 24 hours of runtime, and $571 in API tokens. For builders shipping autonomous agents, the run is less about AGI hype and more about the real cost and engineering pattern of long-horizon tool use.

How the Portal run worked

GPT-6 Astra controlled Portal through the Model Context Protocol (MCP) and a modified SourcePauseTool. The game stayed paused while the model analyzed screenshots and player state. Once it generated an input sequence, the tool unpaused and executed the actions. This think-pause cycle explains why the raw VOD streams add up to roughly 24 hours while the edited highlight reel is only 2 hours.

The entire run was conducted by an enthusiast named CozyBlaze. The model had to map 3D spaces, understand puzzle logic, and plan solutions without human guidance. The final tally: 3,336 tool calls and an API cost of $571.18, covered by a $200 Codex Pro subscription.

What the Portal run means for agent builders

For teams building autonomous agents, the experiment demonstrates a few practical points. First, MCP and SourcePauseTool provide a reusable pattern: pause the environment while the model reasons, then execute the action. Second, the 3,336 tool calls show how quickly an agent can consume tokens when navigating a complex task. The cost of $571 for a single game run scales unpredictably with task length and model capability.

OpenAI positions GPT-6 Astra as state-of-the-art in computer use, browsing, software engineering, cybersecurity, and science. President Greg Brockman called it a "generational leap" and welcomed the AGI era, while Nvidia CEO Jensen Huang similarly claimed AGI had arrived. But builder takeaway is narrower: Astra can handle sustained, multi-step tasks with tool orchestration, but the costs and failure modes are still being discovered.

Practical implications for your stack

If you're building agents that control external tools, the Portal setup is instructive. The code is available on GitHub, so you can inspect the MCP integration and SourcePauseTool modifications. The pattern of pausing the environment to reduce latency on model reasoning is generalizable to any game engine or simulation environment. For production agents, the cost of 3,336 tool calls and $571 for a 24-hour run highlights the need for strict token budgets, retry limits, and cost monitoring. A single agent session could eat a month's API budget if left unchecked.

Why this isn't a benchmark or AGI proof

CozyBlaze themselves admitted there are still many problems to solve and said the run shouldn't be viewed as a benchmark. The experiment used a single model configuration, a specific tool setup, and one game. It doesn't prove GPT-6 Astra can generalize across games or adapt to novel puzzles without prompt engineering. Independent evaluations like the ARC Prize have scored Astra below OpenAI's own AGI bar. Treat the Portal run as a demonstration of capability under ideal conditions, not a signal that AGI is here.

The important part for builders: a hobbyist got GPT-6 Astra to autonomously solve a 3D puzzle game with existing protocols and some custom tooling. That's progress worth watching, but the cost and fragility are equally instructive.