Anthropic, a US-based artificial intelligence (AI) company, has trained its newest model, Claude 3.7 Sonnet, to play a modded version of the 1996 Game Boy classic 'Pokémon Red'.

But this isn't just a nostalgic party trick: the AI's ability to battle gym leaders and navigate pixelated forests could potentially revolutionise how businesses use AI for real-world tasks like coding and problem-solving.

Claude 3.7 Sonnet, unveiled this week, is being hailed by Anthropic as the company's "most intelligent model yet" – a hybrid system that can deliver instant answers or switch to "extended thinking" mode for complex challenges, as per an official blog post. 

Claude entering Brock's gym. You can notice how it's thinking about its steps on the left screen.

Imagine a brain that can either blurt out quick replies or settle into deep concentration, complete with visible thought processes. To prove its mettle, researchers let it loose on Pokémon Red, equipping it with pixel-recognition skills, button controls, and enough memory to sustain 35,000 in-game actions.  

The results? While earlier Claude versions got hilariously stuck in the starting house, Claude 3.7 Sonnet evolved into a proper Pokémon trainer. It battled three gym leaders, collected badges, and even developed strategies like questioning its own assumptions – a skill that translates to real-world coding and engineering tasks. Anthropic cheekily noted the AI became "super effective" at problem-solving, referencing Pokémon's type-matchup battle system.  

Claude entering Brock's gym. You can notice how it's thinking about its steps on the left screen.

But why Pokémon? The retro game – where players explore a world to catch and train creatures – turns out to be just about the perfect test for AI persistence. Claude had to navigate mazes, manage inventory, and make long-term decisions without human guidance. However, the real magic lies in Claude's dual nature. 

Through Anthropic's API, developers can now tell the AI to "think longer" for tough jobs, setting custom token budgets to balance speed and precision. Early adopters are already using it to build websites, debug code, and handle full-stack updates – with design platform Canva praising its "production-ready code with superior design taste", states Anthropic's blog. There's even a new terminal tool, Claude Code, that lets developers delegate engineering tasks directly from their keyboards.  

Claude facing off against a Gym Trainer in Brock's gym.

While Claude's Pokémon skills are charming, Anthropic insists this isn't just fun and games. The same focus that helped it beat Gym Leader Surge could help businesses automate complex workflows or tackle open-ended projects. As for what's next? We're secretly hoping Claude 3.8 will continue its journey to be the very best that no one ever was – but for now, it's busy writing your company's next app! Gotta debug 'em all, indeed.

You can watch the ongoing Claude Plays Pokémon livestream here. Note that the current playthrough is different from the official one conducted by Anthropic.