OpenAI’s latest large language model (LLM), GPT-6 Astra, is making waves for its ability to master a variety of games, including three-dimensional (3D) puzzle games and Pokémon, solely by observing the game screen without human intervention.
Industry sources reported on Tuesday that game creator CozyBlaze announced via X on September 5 (local time) that GPT-6 Astra had successfully completed the 3D puzzle game ‘Portal’ from start to finish without any human input.
Astra, unveiled by OpenAI on September 3, is their next-generation artificial intelligence (AI) model. It reportedly solved puzzles by analyzing only the game screen, much like a human player, without accessing internal sources such as maps or game information.
Using AI models to tackle games is a common method for assessing their performance. OpenAI’s competitor, Anthropic, recently demonstrated its latest model, Claude, completing a game using its new iteration, Fable 5. Fable 5 finished the 1996 game Pokémon FireRed in about 50 hours, relying solely on the game screen without maps or additional game-related information. GPT-6 Astra has shown even more advanced capabilities, with early access users reporting that it cleared Pokémon FireRed in as little as 18 hours and 12 minutes without any specific instructions or hints.
Astra has also demonstrated its ability to independently learn and apply game control methods in various titles, including Slay the Spire 2, Minecraft, and web browser battle games, without any external guidance or programming.
These gaming achievements suggest that Astra has significantly improved reasoning abilities compared to existing AI models, showcasing high problem-solving capacity without the need for sophisticated external frameworks.
In fact, Astra achieved an impressive score of up to 99.9% in the abstract reasoning evaluation benchmark ARC-AGI-3, which tests an AI model’s ability to identify goals and solve problems in unfamiliar environments without predefined rules.
While its performance dropped to 62.7% when using a standard testing framework, Astra still significantly outperformed its competitor, Claude Opus 5, which scored only 30.2%.
However, OpenAI acknowledges that there are still limitations in claiming true artificial general intelligence (AGI). The computational cost for Astra to complete Portal was 571 USD. While the gameplay footage lasted two hours, the total time to complete the game, including reasoning, extended to 24 hours. CozyBlaze commented that OpenAI set a goal in 2016 to develop a single agent capable of mastering various games, but it didn’t expect such rapid progress. Although this game completion isn’t a comprehensive benchmark, it’s remarkable to see a general AI agent autonomously explore and solve a 3D puzzle game.