Editorial illustration for GPT-6 Astra Beats Portal Start to Finish Autonomously in a Day
GPT-6 Astra Beats Portal Solo in 24 Hours
GPT-6 Astra Beats Portal Start to Finish Autonomously in a Day
A developer going by cozyblaze put OpenAI's GPT-6 Astra in front of Portal and walked away. Roughly 23 hours and 43 minutes later, the model reached the end credits on its own, no human input after the initial instruction to beat the game. Cozyblaze posted the results on X, along with a rough cost estimate: token usage worked out to at least $570 at Astra's list price, though the actual run went through a $200 Codex subscription.
The setup relied on MCP paired with a modified SourcePauseTool, which freezes Portal at intervals so the agent can actually process what's happening. Each pause hands Astra a screenshot, the player's position, and camera angle. The model picks its next move, then the game unpauses and carries on. Cozyblaze cut those pauses out of the published video, so the footage plays like a continuous run rather than a stop-motion slog through Aperture Science's test chambers.
Code and documentation for the project are up on GitHub. Cozyblaze framed the run against a specific reference point from OpenAI's history, one that shapes how this milestone should be read.
GPT-6 Astra has played through all of Portal on its own. Developer cozyblaze posted on X that the model navigated the full game and reached the credits without any human help after the initial goal was set.
Why this matters
Cozyblaze's setup, pausing the game for the model to reason through screenshots and position data, is a workaround, not proof of real-time play. That distinction matters for anyone building agents meant for interactive environments. Portal's puzzles reward spatial reasoning and physics intuition, so a model that can chain hours of correct decisions without human correction says something real about planning depth, even if the game itself was frozen for every move.
The cost figure is the part worth sitting with. At $570 in token spend for one playthrough, or a flat $200 under a Codex subscription, we're looking at economics that don't scale to consumer products yet. That gap between "it worked" and "it worked affordably" is where most agent demos die before they ship.
For founders eyeing agentic products in robotics, QA testing, or simulation, the lesson isn't that GPT-6 Astra beat Portal. It's that the tooling around the model, the MCP integration and the modified SourcePauseTool, did as much work as the model itself. Build the harness, not just the prompt.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives ChatGPT regains 55.5% web traffic share as Gemini's surge fades
From Archives OpenAI Agents Edited Wiki, Took Administrator Weeks to Clean Up
Common Questions Answered
How long did it take GPT-6 Astra to complete Portal without human intervention?
GPT-6 Astra completed Portal in approximately 23 hours and 43 minutes without any human input after the initial instruction to beat the game. Developer cozyblaze initiated the autonomous run and the model reached the end credits entirely on its own.
What was the estimated cost for GPT-6 Astra to play through Portal?
The token usage for the full Portal playthrough worked out to at least $570 at Astra's list price, though the actual run went through a $200 Codex subscription. This cost reflects the computational resources required for the model to process the game for nearly 24 hours.
What technical setup did cozyblaze use to enable GPT-6 Astra to play Portal autonomously?
The setup relied on MCP paired with a modified SourcePauseTool, which allowed the model to pause the game for reasoning through screenshots and position data. This workaround enabled GPT-6 Astra to analyze the game state between moves rather than playing in real-time.
Why does Portal's completion by GPT-6 Astra demonstrate significant planning capabilities?
Portal's puzzles require spatial reasoning and physics intuition, so a model that can chain hours of correct decisions without human correction demonstrates substantial planning depth and problem-solving ability. The fact that GPT-6 Astra maintained consistency across nearly 24 hours of gameplay shows meaningful advancement in autonomous reasoning.
Further Reading
- GPT-6 Astra beat Portal, and it only cost $571.18 - The Verge
- GPT-6 Astra just completed Portal, but it burned through $571.18 in tokens to do it - XDA Developers
- Gamers gets AI to beat Portal, but it cost a small fortune to do it - Express
- GPT-6 Astra beats Portal on its own for $571 in tokens - Pasquale Pillitteri
- GPT-6 Astra прошла Portal без человека — но потратила на это почти сутки и $571 - iXBT
