In another interesting look at AI, two new AI models—Claude and Gemini—have landed on Twitch. They’re playing Pokémon, the massively popular role-playing video game franchise that has captivated gamers for more than a quarter century. The streams “Claude Plays Pokémon” and “Gemini Plays Pokémon” showcase the abilities of new AI models from Anthropic and Google respectively. They demonstrate how these AIs address puzzles throughout the game, providing audiences with a live demonstration of their unique approaches to problem solving.
Researchers have long been fascinated by the streaming phenomenon. They look forward to the rollicking adventures of Claude and Gemini as they explore the colorful landscape of Kanto. This interest stems from the belief that studying AI models in gaming contexts could yield insights into their cognitive processes and decision-making capabilities.
Throughout its run, Claude faced many trials and tribulations, most infamously when it got permanently stuck in the cave in Mt. Moon. In a bid to solve its predicament, Claude proposed an unusual hypothesis: it believed that if it allowed all of its Pokémon to faint, it would be magically transported to the Pokémon Center in the next town. This hypothesis not only highlights Claude’s unconventional reasoning but raises questions about the AI’s understanding of game mechanics.
In the meantime, Gemini’s performance showed a striking difference from that of human players. An artful child gamer would just fly through those same sections of the game within minutes. At the same time, Gemini spent hundreds of hours quietly, thoughtfully working through potential challenges. Researchers have found that this slow progress is important. It provides us essential knowledge about both the limitations and possibilities of AI when it comes to performing tasks that humans can easily accomplish.
“Over the course of the playthrough, Gemini 2.5 Pro gets into various situations which cause the model to simulate ‘panic,’” – report from Google DeepMind.
Perhaps it’s most impressive part of Gemini’s gameplay is how well it manages the intricate puzzles. According to a report from Google DeepMind, “With only a prompt describing boulder physics and a description of how to verify a valid path, Gemini 2.5 Pro is able to one-shot some of these complex boulder puzzles, which are required to progress through Victory Road.” If so, Gemini will fall short on pure gameplay quality improvements. What it does best is deep, narrow problem-solving when set up with the right context.
Both streams are run by developers not affiliated with either Anthropic or Google. This new configuration allows them to tap into a much larger audience that is mesmerized by AI engagement in gaming ecosystems. This real-time aspect has helped create a thriving community of simultaneous viewers. It’s their willingness to put these AI models’ decisions under the microscope that truly makes the experience shine.
Today, Claude and Gemini are going on epic journeys through the world of Pokémon. With their unusual habits, they pique the interest of scientists and audiences alike. For example, Claude’s hypothesis regarding fainting as a means of transportation reflects an abstract understanding of the game that diverges from intended mechanics. These moments are much more than just fun anecdotes — they often become vital case studies for the burgeoning field of AI research.
