GPT-6 Astra beat Portal start to finish without human help in under 24 hours
GPT-6 Astra autonomously completed the full Portal game in approximately 23 hours and 43 minutes without human intervention after the initial goal was set. The agent controls the game via MCP and a modified SourcePauseTool that pauses gameplay while the model processes visual and positional data to select inputs. Token usage for the run amounts to at least $570 at Astra's list price, though the developer utilized a $200 Codex subscription. The achievement is framed as a practical step toward Ope
Analysis
TL;DR
- GPT-6 Astra autonomously completed the full Portal game in approximately 23 hours and 43 minutes without human intervention after the initial goal was set.
- The agent controls the game via MCP and a modified SourcePauseTool that pauses gameplay while the model processes visual and positional data to select inputs.
- Token usage for the run amounts to at least $570 at Astra's list price, though the developer utilized a $200 Codex subscription.
- The achievement is framed as a practical step toward OpenAI's 2016 goal of solving multiple games with a single generalist agent.
- Developer cozyblaze described GPT-6 Astra as "the worst model we'll ever get," highlighting both the current capability and the expectation of future improvements.
Why It Matters
This demonstration shows that large language models can now execute long-horizon, multi-step tasks in complex interactive environments without human assistance, a significant milestone for autonomous agent development. It provides a concrete reference architecture for integrating AI agents with legacy software via tool-use protocols like MCP, which is directly applicable to enterprise automation and simulation-based training. The project also grounds the abstract goal of generalist AI agents in a tangible, reproducible experiment that the community can build upon.
Technical Details
- Control Loop & Tooling: The agent uses MCP (Model Context Protocol) to interface with a modified SourcePauseTool, which pauses the Source engine game while the model reasons, then resumes play after input selection.
- Perception & State: During each pause, the model receives screenshots, the player's current position, and camera angle as input to decide the next set of controls.
- Performance & Cost: The full run took 23 hours 43 minutes; token consumption is estimated at a minimum cost of $570 at list price, with the developer noting a $200 Codex subscription was used.
- Open Source: Code and documentation are publicly available on GitHub, enabling replication and extension by other researchers.
- Architectural Context: The system operates as a single agent without task-specific fine-tuning for Portal, aligning with the generalist agent paradigm.
Industry Insight
- Agent Architecture Validation: The successful use of MCP and pause-based tool integration provides a proven pattern for building agents that interact with non-API, stateful environments, accelerating development in robotics and legacy system automation.
- Cost vs. Capability Trade-offs: The token cost highlights the current economic barrier to long-running autonomous tasks; practitioners should anticipate and plan for inference expenses when deploying similar agents in production.
- Strategic Milestone: Beating a full commercial game autonomously serves as a credible benchmark for generalist agent progress, likely increasing investment and research focus on multi-step reasoning in interactive simulations.
Disclaimer: The above content is generated by AI and is for reference only.