A model connected to a game
Anthropic's February 24, 2025 research description included an experiment in which Claude 3.7 Sonnet played Pokémon Red. The setup supplied screen pixels, a basic memory mechanism and callable controls for navigating and pressing buttons. That supporting system allowed the interaction to continue beyond a single conversational context.
The company reported that the model defeated three Gym Leaders. That is a bounded progress claim, not a statement that it finished the entire game. It also means the demonstration should be understood as a model working with a particular memory and control arrangement, rather than an ordinary chat response accomplishing the task alone.
How extended thinking fit the launch
In a separate model announcement published the same day, Anthropic described Claude 3.7 Sonnet as supporting both ordinary responses and an extended-thinking mode. It said API users could set a thinking budget, making the amount of reasoning effort a configurable part of using the model.
That launch announcement also introduced Claude Code as a limited research preview for engineering work. The coding product and the game experiment illustrated different applications of tool-using systems; neither should be mistaken for evidence that success in one setting automatically transfers to another. Time, available tools and the surrounding workflow remain part of any meaningful comparison.
What the experiment can show
A game provides observable milestones and repeated opportunities to try a different approach after failure. Those properties make progress inspectable, but a company-run demonstration is not an independent reliability assessment. To reproduce or compare the result, evaluators would need to account for the setup as well as the model. The published account supports a historical description of the experiment without inventing a livestream observation or claiming general-purpose competence from a game score.