Claude 3.7 Sonnet plays Pokemon on Twitch with 'ClaudePlaysPokemon' - watch the slow, easy-to-follow gameplay

AI company Anthropic reported that its inference model ' Claude 3.7 Sonnet ,' announced on February 25, 2025, outperformed OpenAI's o1, o3-mini, and DeepSeek-R1 in benchmark tests. In addition, Anthropic has demonstrated that as part of a benchmark test to demonstrate the performance of Claude 3.7 Sonnet, it played the Game Boy software 'Pokémon Red,' released in 1996, and defeated three gym leaders. To demonstrate the benchmark test using Pokémon, Anthropic has started streaming ' ClaudePlaysPokemon ' on the game commentary platform Twitch.

'Claude 3.7 Sonnet' and 'Claude Code' have appeared, and they succeeded in defeating three 'Pokemon' gym leaders with performance exceeding OpenAI o1 and DeepSeek-R1 - GIGAZINE

Claude 3.7 Sonnet is a model that can 'reason' difficult tasks and can even play Pokemon, Anthropic says. The previous model, Claude 3.5 Sonnet, which did not have this advanced reasoning ability, was unable to leave its home in the first town, Masara Town, but Claude 3.7 Sonnet reached Kuchiba City and succeeded in defeating the third gym leader, Mattis.

The Twitch distribution screen is as follows, with the Claude 3.7 Sonnet inference results on the left, the game screen on the right, and the Pokemon you own at the bottom.

The distribution started at around 0:00 on February 26, 2025, Japan time, from the Otsukimiyama Cave, which is thought to be a continuation of the benchmark data. However, a new game was started again one hour after the start of distribution, and at the time of writing the article, Tokiwa Forest was being challenged.

Claude 3.7 Sonnet recognizes the screen and controls the buttons accordingly, but because it has to make inferences each time, you can't expect smooth gameplay; you'll progress through the game at an incredibly slow pace.

According to the IT news site TechCrunch, the AI was stuck on a rock wall for several hours after the start of the broadcast and was unable to move. In response to this, one user commented, 'Which would win: an AI programmed with thousands of hours of programming or a rock wall?'

Also, after receiving the Pokemon from Professor Oak, there is a scene where you have to talk to NPCs in the lab multiple times, which caused some users to express concern, asking 'Is this okay?' However, there were also comments such as 'Everyone calm down. I went in and out of Professor Oak's lab about 10 times before I figured out how to proceed,' so it seems best to just be patient and watch over the situation.

In addition, Twitch has also previously broadcast a 'stream aiming to clear Pokemon with fish movements,' which took a whopping 3,195 hours (about 133 days) to be inducted into the Hall of Fame.

A delivery aiming to clear Pokemon with fish took 3195 hours to finally achieve Hall of Fame and clear - GIGAZINE