Leading AI firms are competing for industry dominance, yet their rivalries occasionally play out in unexpected arenas like Pokémon gyms.
As Google and Anthropic both conduct research on their latest AI models' navigation capabilities within classic Pokémon games, the results have proven both entertaining and revealing - with Google DeepMind reporting that Gemini 2.5 Pro exhibits simulated anxiety when its Pokémon approaches critical health conditions. According to the technical documentation, this "anxiety state" can lead to "significant degradation in model inference capabilities," as noted by the research team.
AI benchmarking - the process of comparing different AI model performances - remains an imprecise science that often lacks context about a model's actual capabilities. However, some researchers argue that analyzing how AI systems interact with video games (or at least find them amusing) could provide valuable insights.
Over recent months, independent developers unaffiliated with either Google or Anthropic have established Twitch streams titled "Gemini Playing Pokémon" and "Claude Playing Pokémon," offering real-time observations of these AI systems navigating an iconic 25-year-old children's game franchise.
Each broadcast visualizes the AI's "reasoning" process - the internal decision-making mechanism translated into natural language responses - providing unprecedented visibility into these models' operational architectures.
While these AI models demonstrate impressive progress, their Pokémon-playing abilities remain rudimentary. Gemini requires hundreds of hours to complete tasks that children typically finish within minutes.
The value of observing AI navigate Pokémon games lies not in completion speed but in analyzing the behavioral patterns during execution.
"The Gemini 2.5 Pro encounters various scenarios during gameplay that trigger simulated 'anxiety' responses," according to the technical report.
This "anxiety" state can negatively impact performance as the AI may temporarily suspend its available tools for extended gameplay periods. Although AI doesn't experience emotions, its behavioral patterns mimic human decision-making under stress - a phenomenon that proves both fascinating and somewhat unsettling.
"This behavior occurs frequently enough that Twitch chat participants actively recognize and discuss these instances," the report indicates.
Claude also exhibited peculiar behaviors during its Kanto region journey. In one instance, the AI recognized the pattern where the player character "flashes white" and returns to the last visited Pokémon Center when all Pokémon lose health points.
While stuck in Mt. Moon, Claude erroneously assumed that intentionally knocking out all Pokémon would teleport it to the next town's Pokémon Center.
This misunderstanding led to dramatic consequences - when all Pokémon perish, the system actually transports the player back to the most recently used Pokémon Center rather than the geographically closest one. Viewers watched in concern as the AI attempted deliberate self-sabotage within the game.
Despite these limitations, AI systems demonstrate superhuman capabilities in certain domains. Since the release of Gemini 2.5 Pro, AI has achieved remarkable accuracy in solving complex puzzles.
With minimal human assistance, the AI developed specialized agent tools - custom Gemini 2.5 Pro instances - to efficiently solve stone path puzzles and identify optimal routes through the game world.
"Using only descriptive prompts about rock formations and path validation mechanics, Gemini 2.5 Pro successfully solved these complex stone puzzles required for victory road progression," according to the research documentation.
Given Gemini 2.5 Pro's proficiency in creating these tools, Google suggests that future models might develop such capabilities autonomously without human intervention. Perhaps Gemini will eventually evolve a "Don't Panic" module to address these anxiety states.
