AI Starts Streaming Pokémon Red on Twitch

Anthropic has launched Claude Plays Pokémon on Twitch, where the company’s newest AI bot, Claude 3.7 Sonnet, plays Pokémon Red. The project is an experiment in its own right, designed to demonstrate the capabilities of modern AI technology and how people respond to it.

Image source: Twitch

Researchers have tested AI algorithms in games ranging from Street Fighter to Pictionary, but usually for fun rather than for any real benefit. Meanwhile, Anthropic said Pokémon Red was a benchmark for Claude 3.7 Sonnet, which can effectively “think” through the game’s puzzles.

The Claude 3.7 Sonnet algorithm, like competitors like the OpenAI o3-mini and DeepSeek R1, can “reason” when solving complex problems like playing children’s video games. The previous version of the Claude 3.5 Sonnet algorithm, which lacked the ability to “reason,” failed early on in Pokémon Red. In contrast, the new Claude 3.7 Sonnet has managed to make significant progress.

Image source: Twitch

However, this algorithm also faces difficulties. A few hours after the start of the stream on Twitch, Claude hit a stone wall, which he could not overcome, despite all his efforts. One of the users of the platform even wondered: “Who will win: AI, which was programmed for thousands of hours, or 1 wall?” Eventually, Claude realized that he could bypass the wall.

On the one hand, it’s excruciatingly boring to watch Claude slowly make his way through Pokémon Red, carefully considering every step. But at the same time, the gameplay is strangely captivating. The left side of the screen shows the bot’s “thought process”, and on the right is the game itself.

Image source: Anthropic

For longtime Twitch users, Anthropic’s stream format may seem nostalgic. More than a decade ago, millions of people simultaneously attempted to play Pokémon Red in a first-of-its-kind online social experiment called Twitch Plays Pokémon. Each user could control a character in the game via Twitch chat, which predictably led to chaos in the game. In 2025, however, the platform’s users have become spectators watching an AI algorithm attempt to play a game that many of us could master at age five.