Google's Gemini 2.5 Pro just beat Pokémon Blue

Who would have thought that in 2025, the ultimate proving ground for frontier AI would be a 29-year-old Game Boy cartridge?

After a grueling playthrough spanning hundreds of hours live on Twitch, Google’s Gemini 2.5 Pro Experimental has officially beaten the 1996 classic Pokémon Blue. The moment the credits rolled, Google CEO Sundar Pichai hopped onto X to celebrate: "What a finish! Gemini 2.5 Pro just completed Pokémon Blue!"

This wasn't an official deep-bench research project cooked up inside Mountain View, though. It was the brainchild of Joel Z, a 30-year-old software engineer with no ties to Google, who hooked up the model to run the game largely on its own.

How to teach an LLM to catch 'em all

Pokémon Blue is deceptively brutal for artificial intelligence. It demands open-world navigation, deep tactical planning for turn-based combat, and the patience to solve finicky environmental puzzles. For an AI agent, holding onto long-term goals over hundreds of steps without getting hopelessly lost is a massive headache.

Joel Z solved this by building an emulator harness around mGBA. The setup piped visual screenshots and raw RAM stats—like player coordinates, party health, and map layouts—directly into Gemini 2.5 Pro. To bridge the gap in spatial reasoning, the video feed was overlaid with a visible navigation grid, backed by a text-based map keeping track of where the model had already explored.

From there, Gemini acted as the player, spitting back controller inputs like pressing "A," "B," or steering in a cardinal direction. To prevent the context window from blowing up, the framework automatically condensed earlier gameplay logs roughly every 100 actions.

Even better was the division of labor. Gemini wasn't playing alone; it had help from specialized Gemini sub-agents. A "Pathfinder" agent handled routing through maze-like terrain, while a "Boulder Puzzle Strategist" stepped in to calculate rock-sliding puzzles. The primary Gemini model decided on its own when to call in these specialists, showing a solid grasp of when routine play turned into a complex puzzle.

Help, hacks, and developer interventions

Naturally, the internet wanted to know: how much hand-holding did the model actually get?

Joel Z was candid on stream about his "dev interventions." He occasionally stepped in to fix emulator glitches or rein in bad habits, like the model burning through escape ropes unnecessarily. But he insisted there were no walkthroughs or cheat codes fed into the prompt. The only direct gameplay tip involved telling Gemini to talk to a specific Team Rocket Grunt twice to get the Lift Key—a notorious quirk in the original 1996 code that was later patched in Pokémon Yellow.

This run also highlights the playful rivalry brewing between AI labs. Back in February, Anthropic showed off Claude making strides in Pokémon Red. That led Google AI Studio product lead Logan Kilpatrick to gloat that Gemini had snagged five gym badges while rival models were stuck at three, prompting Pichai to joke about building "Artificial Pokémon Intelligence." Joel Z, however, warned viewers against treating this as a standardized LLM benchmark, noting that Claude and Gemini were running on completely different harnesses with different access to data.

Next on the roadmap: integrating structured note-taking tools, aiming for a completely uninterrupted run, and testing how models like Claude or o3 perform inside the exact same harness.

Why it matters

Beating an old-school RPG might sound like internet spectacle, but we think it's a telling preview of practical agent architecture.

Joel Z's cleverness wasn't in expecting Gemini to be an omniscient gaming savant. It was in building an operational shell that gave the model the right sensory scaffolding: visual grids to compensate for spatial blindness, RAM data for situational awareness, and sub-agents for modular problem solving. That mirrors exactly how enterprise developers will deploy reasoning models in the wild—surrounding the core LLM with specialized tooling to keep it on the rails.

For developers, the takeaway is clear: long-horizon tasks are entirely solvable right now if you treat the model as a workflow orchestrator rather than a solo worker.

At the same time, the project exposes the limits of raw LLM intelligence. Without the external memory resets and the human developer stepping in to un-wedge weird states, Gemini still would have drifted into dead ends. It's not general intelligence, but as a demonstration of coordinated multi-agent planning, it's one of the coolest proofs of concept we've seen all year.