Google’s Gemini has overwhelmed Pokémon Blue (with slightly assist)

Google’s costliest AI mannequin appears to have crossed a serious milestone: Beating a 29-year-old online game.

Final night time, Google CEO Sundar Pichai posted triumphantly on X, “What a end! Gemini 2.5 Professional simply accomplished Pokémon Blue!”

To be clear, the Gemini Performs Pokemon livestream was created by (in his personal phrases) “a 30 yr previous software program engineer unaffiliated with Google” who goes by Joel Z. However Google executives have been cheering the trouble on.

For instance, Logan Kilpatrick, the product lead for Google AI Studio, posted final month that Gemini was “making nice progress at finishing Pokémon” and had “earned its fifth badge (subsequent finest mannequin solely has 3 up to now, although with a distinct agent harness),” main Pichai to joke, “We’re engaged on API, Synthetic Pokémon Intelligence:)”

Why Pokémon? Again in February, Anthropic highlighted progress that its Claude AI fashions had been making in “Pokémon Purple,” writing that Claude’s “prolonged considering and agent coaching” offers it “a serious enhance” on “extra sudden” duties, like enjoying a basic sport. (“Pokémon Purple” and “Blue” are totally different variations of a GameBoy title first launched in 1996 and tied to the long-running Pokémon franchise). There’s even a Claude Performs Pokemon Twitch channel that Joel Z cited as an inspiration.

Regardless of its progress, Claude doesn’t seem to have overwhelmed “Pokémon Purple” but. Does that imply Gemini is objectively higher on the sport? On his Twitch web page, Joel Z urged viewers, “Please don’t think about this a benchmark for a way effectively an LLM can play Pokemon. You possibly can’t actually make direct comparisons — Gemini and Claude have totally different instruments and obtain totally different data.”

And each AI fashions need assistance to play the sport — that’s the place the aforementioned agent harnesses are available, offering the fashions with sport screenshots overlaid with further data, permitting the mannequin to determine reply (which can contain calling specialised brokers), after which urgent the button that corresponds with the AI’s instruction.

Joel Z acknowledged that there have been different “dev interventions” to assist Gemini full the sport, however insisted that it’s not dishonest.

“My interventions enhance Gemini’s general decision-making and reasoning skills,” he says. “I don’t give particular hints — there aren’t any walkthroughs or direct directions for explicit challenges like Mt. Moon. The one factor that comes even shut is letting Gemini know that it wants to speak to a Rocket Grunt twice to acquire the Raise Key, which was a bug that was later fastened in Pokemon Yellow.”

Plus, he mentioned, “Gemini Performs Pokémon remains to be actively being developed, and the framework continues to evolve.”