OCOPD Autonomous Breakthrough in Multi-Hour Computer Interaction
GPT-6 Astra has completed Valve’s classic puzzle game Portal from start to finish in under 24 hours without human assistance during gameplay. Conducted by developer cozyblaze, the experiment marks a notable demonstration of computer-use AI systems handling complex, long-horizon visual and spatial challenges. While human configuration was necessary to set up the execution environment, tool interfaces, and initial prompt, the system received zero human interventions or manual corrections once the playthrough commenced.
The complete codebase and setup documentation have been published on GitHub, with initial reporting from The Decoder and additional analysis by The Verge. Reflecting on the system's rapid progression, cozyblaze noted that Astra represents "the worst model we'll ever get," emphasizing how quickly base capabilities are expected to advance from this baseline.
Experimental Framework and Cost Breakdown
To evaluate the experiment accurately, it is essential to distinguish initial human setup from runtime execution. The developer designed the outer agentic loop, computer-use tool bindings, and prompt directives prior to launch. Once active, the agent parsed visual feedback from the game screen and generated input actions autonomously.
According to financial metrics recorded during the run, the project incurred a total expense of $571.18. This reported figure encompasses:
- Model API Usage: High-volume token consumption for multimodal visual analysis and reasoning chains.
- Inference Tool Overhead: Frequent API calls to translate model outputs into real-time keyboard and mouse commands.
- Hosting and Compute Costs: Infrastructure expenses required to keep the environment, local wrappers, and capture pipelines operational over the 24-hour period.
By operating continuously for nearly a full day, the test demonstrated the model's capacity to maintain spatial orientation, recover from mechanical errors, and retain state across hundreds of sequential puzzle steps.
Implications for Enterprise Computer-Use Models
Solving Portal serves as a demanding testbed for computer-use agents because 3D navigation requires continuous real-time planning, momentum awareness, and cause-and-effect reasoning. Unlike synthetic static benchmarks, interactive virtual environments force models to contend with unexpected failure states without resetting the overarching goal.
This continuous execution holds direct relevance for software automation beyond gaming:
- Complex GUI Automation: Operating desktop applications that lack native APIs.
- Long-Horizon Workflow Execution: Maintaining goal state across multi-hour software testing or data processing tasks.
- Error Recovery: Independently detecting missteps and adjusting strategies without escalation to human operators.
While a single successful game completion does not imply general intelligence or flawless operational reliability, the sub-$600 run provides a concrete benchmark for developers building autonomous desktop agents. As model efficiency improves and inference costs decline, long-duration agentic workflows are poised to move closer to commercial viability.